No entries match these filters. Clear filters
New thinking metadata in /api/show response and MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.34.3 https://github.com/ollama/ollama.git
$ git checkout v0.34.3
└──▷ USE IT
Check thinking configuration for a cloud model to understand default behavior.
shell
copy
curl http://localhost:11434/api/show -d '{"model": "glm-5.3-flash:cloud"}'
GET /api/show response now includes a thinking object with values and default keys for model thinking controlsSupports Nemotron H vision models via MLX backend on Apple Silicon hardware
2 details in Everything
New ollama://apps URI scheme opens desktop app's Apps page; first-run sign-in flow shared with desktop.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.34.2 https://github.com/ollama/ollama.git
$ git checkout v0.34.2
└──▷ TRY IT
Open the installed models page in the Ollama desktop app from a terminal or browser link on macOS/Windows.
copy
$ open ollama://apps
Adds ollama://apps URI scheme to open the desktop app’s Apps page directly on macOS and Windows Adds first-run setup options (sign in or continue locally) that sync with the desktop app on macOS and Windows
2 details in Everything
Ollama v0.34.0 adds ChatGPT Desktop integration, OpenAI tool search support, and response compaction
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.34.0 https://github.com/ollama/ollama.git
$ git checkout v0.34.0
Adds openai_compatible tool search protocol support for client-side function calling function calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. Adds response compaction feature to reduce context window context window The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it.
2 details in Everything
Ollama v0.33.3 adds image and audio support for gemma4 on MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.33.3 https://github.com/ollama/ollama.git
$ git checkout v0.33.3
└──▷ TRY IT
Run a multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
copy
$ ollama run gemma4
> ![image.png]
> /path/to/audio.mp3
What is in this file? Check how many tokens were served from the context cache during generation.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "gemma4", "prompt": "Hello world"}' | jq '.cached_prompt_tokens'
Adds image and audio modality support for the gemma4 model when running on the MLX engine Reports cached prompt token counts in API responses
2 details in Everything
Ollama v0.33.1 adds structured output structured output Forcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.33.1 https://github.com/ollama/ollama.git
$ git checkout v0.33.1
Updates MLX and llama.cpp backends
1 detail in Everything
Ollama v0.33.0 adds Claude Desktop integration and fixes KV cache KV cache Memory where a model keeps the attention keys and values of tokens it has already processed, so each new token does not recompute them. Its size grows with context length and often limits serving.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.33.0 https://github.com/ollama/ollama.git
$ git checkout v0.33.0
Adds claude_desktop_config.json integration to configure Ollama as a third-party gateway provider for Claude Desktop.
1 detail in Everything
Ollama v0.32.14 adds WebP image transcoding for vision models and tolerates non-leading system messages in Qwen.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.14 https://github.com/ollama/ollama.git
$ git checkout v0.32.14
Automatically transcodes WebP image payloads to PNG before forwarding to llama-server, enabling WebP support for vision model integrations alongside existing JPG and PNG formats. Qwen renderer now passes non-leading system messages through the raw ChatML path instead of returning an HTTP 500, allowing coding clients that inject runtime system prompts mid-conversation to function correctly.
2 details in Everything
Ollama v0.32.13 adds Qwen3.8 developer-instruction support, folding leading system/developer prefixes into a single system turn.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.13 https://github.com/ollama/ollama.git
$ git checkout v0.32.13
Ollama v0.32.12 adds Qwen3.8 27B support with an Apple Silicon MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.12 https://github.com/ollama/ollama.git
$ git checkout v0.32.12
└──▷ TRY IT
Run the Qwen3.8 27B MLX variant on Apple Silicon for high-quality coding agent workloads.
copy
$ ollama run qwen3.8:27b-mlx
Adds qwen3.8:27b model, with performance gains across coding, research, and long-horizon agentic tasks, runnable via ollama run qwen3.8:27b. Adds qwen3.8:27b-mlx variant optimized for maximum performance and output quality on Apple Silicon devices, runnable via ollama run qwen3.8:27b-mlx.
2 details in Everything
Ollama v0.32.11 adds DeepSeek Harness and Muse Code integrations and brings web search to the OpenAI Responses API.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.11 https://github.com/ollama/ollama.git
$ git checkout v0.32.11
└──▷ TRY IT
Launch DeepSeek's open-source agent harness agent harness The software around an AI agent that manages its tools, execution loop, state and limits. It determines how the model's proposed actions are run and checked.
Launch Meta's Muse Code agentic coding CLI backed by a local Ollama model.
copy
$ ollama launch muse
Adds ollama launch dsh to integrate DeepSeek Harness, DeepSeek's open-source agent harness. Adds ollama launch muse to integrate Muse Code, Meta's agentic coding CLI. Adds support for the Nemotron 3.5 prompt layout, selecting the correct parser and renderer from its checkpoint template.
3 details in Everything
api-docs-94d22f0572de
Aug 12, 2026
·
issue 043
Ollama API docs add a push-a-blob endpoint.
└──▷ GET THIS VERSION
copy
$ git clone --branch api-docs-94d22f0572de https://github.com/ollama/ollama.git
$ git checkout api-docs-94d22f0572de
New push a blob API operation documented in the Ollama API reference.
1 detail in Everything
Ollama v0.32.9 adds NVIDIA Nemotron 3.5 Lightning (30B MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. speculative decoding speculative decoding Speeding up generation by letting a small draft model guess several tokens ahead and having the large model check them all at once. Output is unchanged; accepted guesses save time.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.9 https://github.com/ollama/ollama.git
$ git checkout v0.32.9
└──▷ TRY IT
Run the new Nemotron 3.5 Lightning 30B MoE model locally for always-on agent workloads.
copy
$ ollama run nemotron-3.5-lightning
Adds nemotron-3.5-lightning, an NVIDIA 30B mixture-of-experts mixture-of-experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. ollama run nemotron-3.5-lightning. Adds the Nemotron 3 architecture to the MLX backend, including Mamba2/recurrent layers, MoE routing, and quantized quantized Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. NVFP4 NVFP4 NVIDIA's four-bit floating-point format with scaling for small groups of values. It reduces model memory and supports fast calculations on compatible GPUs. Serves Nemotron 3's multi-token prediction head as a built-in self-draft speculator, enabling speculative decoding without a separate draft model.
3 details in Everything
Ollama v0.32.8 adds Muse Glimmer model support across NVIDIA, AMD, and all platforms for local coding agents and personal assistants.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.8 https://github.com/ollama/ollama.git
$ git checkout v0.32.8
└──▷ TRY IT
Run Muse Glimmer locally with Claude Code as a coding agent backend — no cloud API key required.
copy
$ ollama launch claude --model muse-glimmer Use Muse Glimmer with OpenClaw to self-host a personal AI assistant across messaging platforms.
copy
$ ollama launch openclaw --model muse-glimmer
Adds ollama run muse-glimmer to download and run the Muse Glimmer model locally on NVIDIA, AMD, and additional platforms. Adds ollama launch claude --model muse-glimmer to run Muse Glimmer with Claude Code as a local coding agent. Adds ollama launch pi --model muse-glimmer to run Muse Glimmer with the Pi coding agent integration. Adds ollama launch openclaw --model muse-glimmer and ollama launch hermes --model muse-glimmer to power personal assistant frameworks locally.
4 details in Everything
Ollama v0.32.7 adds Muse Glimmer 30B with DFlash and image input on Apple Silicon, plus streaming thinking traces and a ctrl+o trace toggle in the TUI TUI A text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.7 https://github.com/ollama/ollama.git
$ git checkout v0.32.7
└──▷ TRY IT
Run Muse Glimmer locally on Apple Silicon for interactive chat or to test multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
copy
$ ollama run muse-glimmer:30b-mlx Wire Muse Glimmer into Claude Code as a local coding agent on Apple Silicon.
copy
$ ollama launch claude --model muse-glimmer:30b-mlx Launch Muse Glimmer with the OpenClaw personal assistant framework for WhatsApp, Telegram, Slack, or Discord.
copy
$ ollama launch openclaw --model muse-glimmer:30b-mlx
Adds muse-glimmer:30b-mlx model tag, enabling the Muse Glimmer 30B multimodal model via ollama run muse-glimmer:30b-mlx or ollama launch <integration> --model muse-glimmer:30b-mlx. Adds DFlash block-diffusion drafting support for the MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. Adds image input support to the MLX engine (Muse Glimmer on Apple Silicon), enabling multimodal vision workloads locally. Streams live thinking deltas beneath a Thinking ↓ N tokens row in the agent TUI, collapsing to a persistent Thought marker when the assistant begins responding. Expands namespace tool declarations in the OpenAI Responses API, resolving undeclarable namespaced tool calls by unfolding nested tools arrays into namespace-qualified flat function names.
5 details in Everything
Ollama v0.32.6 adds automatic MTP speculative decoding speculative decoding Speeding up generation by letting a small draft model guess several tokens ahead and having the large model check them all at once. Output is unchanged; accepted guesses save time. GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.6 https://github.com/ollama/ollama.git
$ git checkout v0.32.6
└──▷ TRY IT
Stream a chat completion and receive token usage in a separate trailing chunk, matching OpenAI's wire format.
copy
$ curl http://localhost:11434/v1/chat/completions -d '{"model": "gemma4", "stream": true, "stream_options": {"include_usage": true}, "messages": [{"role": "user", "content": "Why is the sky blue?"}]}'
Adds stream_options.include_usage support in /v1/chat/completions so usage is returned in a separate trailing chunk, matching OpenAI's streaming wire format — role now appears only on the first chunk and finish_reason on its own chunk. Qwen3.5 inference on Apple GPUs is faster: the MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. Experimental image generation has been removed; continue using v0.32.5 for image generation support.
3 details in Everything
└──▷ BREAKING ON UPGRADE
! Experimental image generation is no longer available in v0.32.6; the feature has been removed and requires staying on v0.32.5.
Ollama v0.32.4 adds Laguna MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.4 https://github.com/ollama/ollama.git
$ git checkout v0.32.4
Adds support for the Laguna model family on Apple Silicon via the MLX engine, with Metal memory residency for significantly improved prompt and generation performance on qualifying hardware. Adds permission/approval flow for model-initiated agent skill loading, while preserving direct user slash-skill activation without prompts. Adds /system prompt inspection and on/off toggle commands to the agent TUI TUI A text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. Quantizes draft-model output heads at the requested quantization type when creating speculative-decoding drafts. Supports per-expert quantization format decoding for Qwen3 MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
5 details in Everything
Ollama v0.32.3 adds CUDA CUDA NVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first. ARM64 ARM64 The 64-bit instruction set architecture from ARM Holdings, used in modern mobile chips, Apple Silicon, and cloud server CPUs; tools target it to run natively on those platforms without emulation overhead.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.3 https://github.com/ollama/ollama.git
$ git checkout v0.32.3
Adds CUDA support on Windows ARM64 devices. Adds NVIDIA B200 GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Reduces memory usage on Linux CUDA and ROCm ROCm AMD's open software stack for running AI and compute workloads on its GPUs, the counterpart to NVIDIA's CUDA. GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Adds chat, thinking, and tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. Restores Claude Code Channels support when launching via ollama launch claude.
5 details in Everything
Ollama v0.32.1 adds CWD-aware agent context, actionable web-auth guidance, and MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.1 https://github.com/ollama/ollama.git
$ git checkout v0.32.1
Adds current working directory to the interactive agent's system prompt for better project context. Agent web search and fetch now surface an actionable ollama signin prompt when authentication is required, instead of a generic error. MLX text model loading now respects the OLLAMA_LOAD_TIMEOUT environment variable. Improves Gemma 4 tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
4 details in Everything
Ollama v0.32.0 adds an interactive agent TUI TUI A text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.32.0 https://github.com/ollama/ollama.git
$ git checkout v0.32.0
└──▷ TRY IT
Start the new interactive agent to chat, write code, and delegate work — no model name needed.
Route local model traffic through the ChatGPT desktop app (formerly Codex App ) integration.
copy
$ ollama launch chatgpt
Adds an interactive agent experience: running ollama now launches a built-in agent for chat, coding, web search, and delegating real work. Adds --restore flag to ollama launch chatgpt to return to your usual ChatGPT profile. Simplifies the ollama launch root menu to a curated short list of popular integrations; all others remain accessible via ollama launch <integration>. Warns before launching older/weaker agent models (CodeLlama, Qwen2.5, Qwen2.5-Coder, Llama 3.x, Mistral, StarCoder, base DeepSeek-R1 tags); use --yes to bypass in headless workflows.
4 details in Everything
Ollama v0.31.2 adds flash attention on Pascal GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.31.2 https://github.com/ollama/ollama.git
$ git checkout v0.31.2
└──▷ TRY IT
Launch Claude Code backed by a local model with telemetry and feedback reporting disabled by default (no extra flags needed — privacy-off is now the default).
copy
$ ollama launch claude
Enables flash attention on older NVIDIA Pascal GPUs (compute capability 6.x), unlocking faster inference on GTX 10-series and similar cards. Enables iGPU to offload vision models with padding to fit available memory, broadening multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. Disables Claude Code telemetry, error reporting, feedback commands, and non-essential traffic by default when using ollama launch claude. Updates the MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
4 details in Everything
Gemma 4 on Apple Silicon gets ~90% faster token generation via automatic multi-token prediction (MTP), no config required.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.31.1 https://github.com/ollama/ollama.git
$ git checkout v0.31.1
Enables automatic multi-token prediction (MTP) for Gemma 4 on Apple Silicon, delivering ~90% faster token generation on coding-agent workloads with no configuration or output changes. Adds minimum-version enforcement for ollama launch hermes-desktop, automatically upgrading older Hermes CLI installs (v0.16.0+ required) instead of failing silently. Improves streamed tool-call parsing so closing braces or brackets inside JSON string argument values no longer cause tool calls to be silently dropped as plain text.
3 details in Everything
Ollama v0.30.11 adds auto-install for Claude Code and OpenCode, thinking-capability detection, and improved Codex model-drift detection.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.11 https://github.com/ollama/ollama.git
$ git checkout v0.30.11
└──▷ TRY IT
Let Ollama install Claude Code for you on first use, instead of installing it manually before launching.
copy
$ ollama launch claude Let Ollama install OpenCode for you on first use; on Windows it will run npm install -g opencode-ai@latest automatically.
copy
$ ollama launch opencode
Adds auto-install for Claude Code: running ollama launch claude now prompts to install Claude Code automatically if the binary is missing. Adds auto-install for OpenCode: running ollama launch opencode prompts to install OpenCode when it is not found, using the official install script on macOS/Linux and npm on Windows. Adds thinking-capability detection for OpenCode: ollama launch opencode detects if the selected model supports reasoning and writes the appropriate config, enabling users to toggle thinking via Ctrl+T in OpenCode's TUI TUI A text-based user interface that runs inside a terminal, using characters and ANSI escape codes to render menus, panes, and widgets without a graphical desktop, letting cyber tools ship rich interactivity with no GUI dependency. Adds model-drift detection for Codex App: ollama launch codex-app now detects when the Codex UI has switched to a non-Ollama model and marks the integration as inactive rather than silently mis-routing requests. Adds sm_86 GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. CUDA CUDA NVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first. GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
+4 more
show less
Adds Compute Capability 87 (Jetson Orin) support for CUDA v13 on Jetson devices. Enables the host Vulkan loader on Windows, broadening Vulkan backend compatibility. Improves speculative decoding speculative decoding Speeding up generation by letting a small draft model guess several tokens ahead and having the large model check them all at once. Output is unchanged; accepted guesses save time. MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. Updates the bundled llama.cpp version, bringing in upstream model and performance improvements.
9 details in Everything
Command A and North family models now run on Apple Silicon via the MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.10 https://github.com/ollama/ollama.git
$ git checkout v0.30.10
└──▷ TRY IT
Run a Cohere North model locally on an Apple Silicon Mac using the MLX engine — now supported for the first time.
copy
$ ollama run command-a
Adds MLX engine support for Cohere Command A and North family models (including North-Mini-Code) on Apple Silicon Macs. Updates the llama.cpp backend to build b9672.
2 details in Everything
Ollama v0.30.9 adds Cohere2Moe support and context-limit error reporting for oversized messages.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.9 https://github.com/ollama/ollama.git
$ git checkout v0.30.9
Adds support for the Cohere2Moe model architecture. Returns an error when a single message exceeds the current context window context window The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it.
2 details in Everything
Ollama v0.30.8 decouples prompt caching from context shift and adds a new cache_prompt API field for finer KV cache KV cache Memory where a model keeps the attention keys and values of tokens it has already processed, so each new token does not recompute them. Its size grows with context length and often limits serving.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.8 https://github.com/ollama/ollama.git
$ git checkout v0.30.8
└──▷ TRY IT
Disable prompt caching for a single request without disabling context shifting — useful when you want a clean KV cache for a fresh conversation without changing overflow behaviour.
copy
$ curl http://localhost:11434/api/chat -d '{"model": "llama3.2", "messages": [{"role": "user", "content": "Explain TLS handshake"}], "cache_prompt": false, "stream": false}'
Adds a new top-level cache_prompt field to generate and chat API requests, letting callers independently control KV cache reuse without affecting context-shift (shift) behavior. Improves KV cache reuse by decoupling prompt caching from context shift, so shift: false no longer silently disables prompt caching. Enhances MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. Adds prefill prefill The first phase of inference, where the model reads the whole prompt in parallel before producing its first token. It drives time-to-first-token; decode is the token-by-token phase after it. Improves recurrent model support by returning per-boundary states from gated-delta kernels.
5 details in Everything
Ollama v0.30.7 adds Hermes Desktop launch support and aligns the OpenAI-compatible models list with available tags.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.7 https://github.com/ollama/ollama.git
$ git checkout v0.30.7
└──▷ TRY IT
Launch the Hermes Desktop visual interface alongside your Hermes agent for a GUI GUI A graphical user interface, controls such as windows, buttons and menus. It lets people use software visually rather than entering commands.
copy
$ ollama launch hermes-desktop Override the Hermes config directory (e.g. a shared or non-default location) before launching Hermes Desktop.
copy
$ HERMES_HOME=/opt/hermes-config ollama launch hermes-desktop
Adds ollama launch hermes-desktop to start a native desktop interface for the Hermes agent with conversation, integration, and messaging-app management. Supports HERMES_HOME environment variable to override the Hermes config directory on any platform. Uses the native Windows %LOCALAPPDATA%\hermes config path for Hermes Desktop, ensuring Ollama-written config is correctly picked up on Windows. Aligns the OpenAI-compatible /v1/models endpoint model identifiers with those returned by /api/tags.
4 details in Everything
Ollama v0.30.6 adds Oh My Pi (omp) coding-agent integration and QAT-optimized Gemma 4 model weights model weights The numerical values a model learns during training. Loading them into the model's architecture restores its learned behavior for inference or further training.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.6 https://github.com/ollama/ollama.git
$ git checkout v0.30.6
└──▷ TRY IT
Launch the Oh My Pi AI coding agent from the terminal, automatically configured with Ollama as the backend provider.
Run the memory-efficient QAT variant of Gemma 4 31B locally, using significantly less VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory. quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
copy
$ ollama run gemma4:31b-it-qat
Adds ollama launch omp integration with Oh My Pi, an AI coding agent with IDE integration that runs in the terminal. Adds Gemma 4 QAT model tags (gemma4:*-qat) with Quantization-Aware Training for dramatically reduced memory requirements on-device. Enables NVFP4 NVFP4 NVIDIA's four-bit floating-point format with scaling for small groups of values. It reduces model memory and supports fast calculations on compatible GPUs. MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
3 details in Everything
Ollama v0.30.5 adds Hermes Desktop launch support and native Windows Hermes install via PowerShell.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.5 https://github.com/ollama/ollama.git
$ git checkout v0.30.5
└──▷ TRY IT
Launch the Hermes Desktop AI agent app directly from the CLI — installs on first run, skips rebuild if already installed.
copy
$ ollama launch hermes-desktop Install and launch the Hermes agent on a native Windows machine without needing WSL WSL Windows Subsystem for Linux, a Microsoft-built compatibility layer that lets a Linux environment run directly on Windows without a virtual machine, giving cyber tools access to Linux binaries, filesystems, and syscalls on a Windows host.
copy
$ ollama launch hermes
Adds ollama launch hermes-desktop to launch the Hermes Desktop app, skipping rebuild when a packaged desktop app is already installed. Enables ollama launch hermes on native Windows installs via the Hermes PowerShell installer.
2 details in Everything
Ollama v0.30.4 adds Nemotron-3-Ultra, Metal GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.4 https://github.com/ollama/ollama.git
$ git checkout v0.30.4
└──▷ TRY IT
Run NVIDIA's new high-throughput reasoning model for agentic workflows.
copy
$ ollama run nemotron-3-ultra Build an MLX-based model that enforces a minimum Ollama version via the REQUIRES directive in its Modelfile.
copy
$ ollama create --experimental my-mlx-model -f ./Modelfile
Adds NVIDIA Nemotron-3-Ultra model, built for high-throughput reasoning and long-running agent workflows (ollama run nemotron-3-ultra). Enables Metal GPU offload for multimodal (mmproj) models on Apple Silicon, accelerating vision/multimodal inference on supported Macs. Supports REQUIRES directive in Modelfiles for MLX-based models when using ollama create --experimental. Adds automatic migration for ollama launch pi users from the legacy @mariozechner/pi-coding-agent package to the official @earendil-works/pi-coding-agent, preserving npm install prefix.
4 details in Everything
Ollama v0.30.3 adds support for Gemma 4 12B, Google's multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.3 https://github.com/ollama/ollama.git
$ git checkout v0.30.3
└──▷ TRY IT
Pull and interactively chat with Gemma 4 12B locally for multimodal reasoning tasks.
copy
$ ollama run gemma4:12b
Adds Gemma 4 12B model support — a multimodal model combining efficiency and advanced reasoning that runs on consumer laptops.
1 detail in Everything
ollama launch adds Qwen Code + Cline CLI guidance, Laguna architecture support, and Radeon 8060S GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.2 https://github.com/ollama/ollama.git
$ git checkout v0.30.2
└──▷ TRY IT
Launch a Qwen Code coding session — Ollama will prompt you to install Cline if it isn't already present.
copy
$ ollama launch qwen-code Launch Codex in an isolated configuration so it won't interfere with your existing Codex settings.
copy
$ ollama launch codex
Adds Qwen Code support to ollama launch, with guided Cline CLI installation when the CLI is missing. Supports Poolside's Laguna model architecture in the llama.cpp backend. Enables Radeon 8060S integrated GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Includes template details in logs to aid troubleshooting of model prompt formatting.
4 details in Everything
Ollama v0.24 launches Codex App support with built-in browser, review mode, and reworked MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.24.0 https://github.com/ollama/ollama.git
$ git checkout v0.24.0
└──▷ TRY IT
Launch the Codex App to work on coding tasks with parallel threads, worktree support, and git integration.
copy
$ ollama launch codex-app Roll back to a previous Codex App configuration after an upgrade.
copy
$ ollama launch codex-app --restore
Adds ollama launch codex-app to run OpenAI's Codex desktop experience with parallel thread support, built-in worktree management, and git functionality. Codex App includes a built-in browser for loading local servers and annotating pages directly to request code changes. Codex App includes a Review mode for leaving inline comments and iterating on code without leaving the workspace. Supports ollama launch codex-app --restore to revert to a previous Codex App configuration. Reworked MLX sampler improves generation quality on Apple Silicon.
5 details in Everything
ollama launch opencode now accepts image inputs when using vision models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.23.4 https://github.com/ollama/ollama.git
$ git checkout v0.23.4
└──▷ TRY IT
Use a vision model interactively with image inputs through the opencode launcher.
copy
$ ollama launch opencode
Ollama v0.30.0 adds llama.cpp engine support, broader GGUF GGUF The single-file model format used by llama.cpp and Ollama, holding the weights, usually quantized, together with the tokenizer and settings needed to run the model locally.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.30.0 https://github.com/ollama/ollama.git
$ git checkout v0.30.0
Adds llama.cpp engine as a backend, augmenting the existing MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. Supports GGUF-based models from Hugging Face directly, including user fine-tuned models. Delivers faster inference performance on NVIDIA hardware.
3 details in Everything
└──▷ BREAKING ON UPGRADE
! nomic-embed-text now converts inputs to lowercase per the model card; prior Ollama versions incorrectly preserved mixed case, so embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
Ollama v0.23.2 caches /api/show for ~6.7x faster model-load latency and refines launch integration management.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.23.2 https://github.com/ollama/ollama.git
$ git checkout v0.23.2
└──▷ TRY IT
Restore Claude Desktop to its default state after Ollama previously modified it via the launch integration.
copy
$ ollama launch claude-desktop --restore
Caches /api/show API responses, cutting median latency by ~6.7x and accelerating integrations such as VS Code. Adds ollama launch claude-desktop --restore command to return Claude Desktop to its normal state after integration removal. Improves backup workflow when managing ollama launch integrations. Delivers a cleaner image-generation layout in the MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
4 details in Everything
└──▷ BREAKING ON UPGRADE
! ollama launch no longer includes Claude Desktop; the third-party integration has been removed because it was limited to Anthropic models.
Ollama v0.23.1 adds Gemma 4 MTP speculative decoding speculative decoding Speeding up generation by letting a small draft model guess several tokens ahead and having the large model check them all at once. Output is unchanged; accepted guesses save time.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.23.1 https://github.com/ollama/ollama.git
$ git checkout v0.23.1
└──▷ TRY IT
Run Gemma 4 31B with speculative decoding on a Mac to get 2x+ faster completions on coding tasks.
copy
$ ollama run gemma4:31b-coding-mtp-bf16
Ollama v0.23.0 adds ollama launch claude-desktop support and server-driven featured model recommendations.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.23.0 https://github.com/ollama/ollama.git
$ git checkout v0.23.0
└──▷ TRY IT
Launch Claude Desktop with Cowork and Code support without leaving your Ollama workflow.
copy
$ ollama launch claude-desktop Use Claude Code in the terminal via the existing CLI entry point.
copy
$ ollama launch claude
Adds ollama launch claude-desktop to launch Claude Desktop (including Claude Cowork and Claude Code) directly from Ollama. Surfaces featured models in the Ollama app via server-driven recommendations.
2 details in Everything
Ollama v0.22.1 updates the Gemma 4 renderer for thinking and tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.22.1 https://github.com/ollama/ollama.git
$ git checkout v0.22.1
Updated Gemma 4 renderer with improved thinking and tool calling support. Model recommendations now refresh independently, without requiring an Ollama update. Aligns the desktop app's launch page with ollama launch integrations.
3 details in Everything
Ollama v0.22.0 adds NVIDIA Nemotron 3 Omni and Poolside Laguna XS.2 to the model library.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.22.0 https://github.com/ollama/ollama.git
$ git checkout v0.22.0
└──▷ TRY IT
Pull and run Poolside's open-weight coding model to get inline code completions or explanations from the CLI.
copy
$ ollama run laguna-xs.2 Pull and run NVIDIA's Nemotron 3 Omni model for multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
copy
$ ollama run nemotron3
Adds NVIDIA Nemotron 3 Omni multimodal model via ollama run nemotron3. Adds Poolside's Laguna XS.2, an open-weight coding-focused model, via ollama run laguna-xs.2.
2 details in Everything
Ollama v0.21.3-rc0 adds max as a think value and maps OpenAI reasoning effort to think parameter.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.21.3-rc0 https://github.com/ollama/ollama.git
$ git checkout v0.21.3-rc0
└──▷ TRY IT
Use maximum thinking depth when generating a response via the Ollama API.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "qwq", "prompt": "Solve this step by step: ...", "think": "max"}'
Accepts "max" as a valid value for the think parameter, enabling maximum reasoning depth in a single flag. Maps OpenAI Responses API reasoning_effort field to Ollama's think parameter, enabling OpenAI-compatible clients to control thinking mode.
2 details in Everything
Ollama v0.21.2 bundles web search in OpenClaw and stabilizes the ollama launch onboarding flow.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.21.2 https://github.com/ollama/ollama.git
$ git checkout v0.21.2
└──▷ TRY IT
Launch the OpenClaw interface to get web search-augmented model responses immediately, with no additional plugin configuration.
OpenClaw integration now includes Ollama's web search plugin out of the box, enabling web-augmented queries without manual plugin setup.
1 detail in Everything
Ollama v0.21.1 adds Kimi CLI integration, logprobs support in MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.21.1 https://github.com/ollama/ollama.git
$ git checkout v0.21.1
└──▷ TRY IT
Run Kimi K2.6 as an agentic assistant for long-horizon tasks directly from the CLI.
copy
$ ollama launch kimi --model kimi-k2.6:cloud
New ollama launch command installs and runs the Kimi CLI, enabling long-horizon agentic execution via a multi-agent system with models like kimi-k2.6:cloud. MLX runner now supports logprobs for compatible models. Faster MLX sampling via fused top-P and top-K in a single sort pass, with repeat penalties applied in the sampler. Improved MLX prompt tokenization by moving tokenization into request handler goroutines. GLM4 MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
5 details in Everything
Ollama v0.21.0 adds Hermes Agent, GitHub Copilot CLI integration, and Gemma 4 on Apple Silicon MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.21.0 https://github.com/ollama/ollama.git
$ git checkout v0.21.0
└──▷ TRY IT
Launch the Hermes Agent to get an AI assistant that learns your workflows and builds skills over time.
copy
$ ollama launch hermes
Adds ollama launch hermes to run the Hermes Agent, which automatically creates skills to better serve research and engineering workflows. Adds GitHub Copilot CLI as a supported integration in ollama launch, configurable alongside other coding agents in a single command. Supports running Gemma 4 via the MLX backend on Apple Silicon, including a text-only MLX runtime. Adds mixed-precision quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. Expands MLX backend with new op wrappers: Conv2d, Pad, activations, trig, masked SDPA, and RoPE RoPE Rotary position embeddings, how most current models encode the order of tokens. Scaling methods such as YaRN stretch it to extend a model's context window.
+1 more
show less
ollama launch opencode now writes config inline instead of to a separate file, consistent with other integrations.
6 details in Everything
Ollama v0.20.5 adds OpenClaw messaging-channel setup and flash attention for Gemma 4.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.20.5 https://github.com/ollama/ollama.git
$ git checkout v0.20.5
└──▷ TRY IT
Spin up an OpenClaw instance to route WhatsApp, Telegram, or Discord traffic through a local Ollama model.
copy
$ ollama launch openclaw
New ollama launch openclaw command sets up OpenClaw channels, connecting WhatsApp, Telegram, Discord, and other messaging platforms. Enables flash attention for Gemma 4 on compatible GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. ollama launch opencode now auto-detects curl-based OpenCode installs at ~/.opencode/bin.
3 details in Everything
Ollama v0.20.3 adds latest models to the app and improves Gemma 4 tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.20.3 https://github.com/ollama/ollama.git
$ git checkout v0.20.3
Adds latest models to the Ollama App model library. Improves Gemma 4 tool calling capabilities.
2 details in Everything
Ollama v0.20.0 adds Google Gemma 4 model family support across four sizes including MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.20.0 https://github.com/ollama/ollama.git
$ git checkout v0.20.0
└──▷ TRY IT
Run the lightweight Gemma 4 E2B model for fast, resource-efficient local inference.
copy
$ ollama run gemma4:e2b Run the Gemma 4 26B Mixture-of-Experts Mixture-of-Experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
copy
$ ollama run gemma4:26b
Supports Google Gemma 4 models in four sizes: E2B, E4B, 26B (Mixture of Experts with 4B active parameters), and 31B (Dense). Adds SentencePiece-style BPE tokenizer support.
2 details in Everything
Ollama v0.19.0 powers Apple Silicon with MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. ollama launch pi.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.19.0 https://github.com/ollama/ollama.git
$ git checkout v0.19.0
└──▷ TRY IT
Run the Pi assistant with web search enabled to answer questions requiring live information.
Adds MLX-powered inference backend for Apple Silicon, leveraging unified memory architecture for improved performance. Adds web search plugin to ollama launch pi, using Ollama's built-in web search capability. MLX runner now creates periodic snapshots during prompt processing prompt processing The first phase of inference, where the model reads the whole prompt in parallel before producing its first token. It drives time-to-first-token; decode is the token-by-token phase after it.
3 details in Everything
Ollama v0.18.3 adds direct VS Code integration via GitHub Copilot, letting local models be used inside the editor.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.18.3 https://github.com/ollama/ollama.git
$ git checkout v0.18.3
└──▷ HOW TO FIND IT
Use a locally running Ollama model as your Copilot model inside VS Code without leaving the editor.
📍 In VS Code with GitHub Copilot, open the model picker in the Copilot chat panel and select any model served by your local Ollama instance.
Enables VS Code GitHub Copilot integration so any local or cloud Ollama model can be selected directly inside Visual Studio Code.
1 detail in Everything
Ollama v0.18.1 adds web search/fetch for OpenClaw and headless mode for ollama launch.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.18.1 https://github.com/ollama/ollama.git
$ git checkout v0.18.1
└──▷ TRY IT
Run a CI pipeline step that asks a cloud model to review a repository without any interactive prompts.
copy
$ ollama launch claude --model kimi-k2.5:cloud --yes -- -p "how does this repository work?" Add web search to an already-configured OpenClaw instance so models can retrieve live content during sessions.
copy
$ openclaw plugins install @ollama/openclaw-web-search Spin up OpenClaw with Ollama's official auth and model provider to enable web search and fetch in an interactive session.
copy
$ ollama launch openclaw
Adds web search and web fetch plugin for OpenClaw via ollama launch openclaw, enabling local and cloud models to retrieve and extract readable web content. Supports installing the web search capability directly into an existing OpenClaw setup with openclaw plugins install @ollama/openclaw-web-search. Adds non-interactive (headless) mode to ollama launch via --yes flag and required --model argument, enabling automated pipelines, CI/CD security checks, and scripted prompt evaluation. Enables OpenClaw subagent task execution in headless mode for automated multi-step workflows.
4 details in Everything
Ollama v0.18 adds non-interactive launch support, Nemotron-3-Super, ROCm ROCm AMD's open software stack for running AI and compute workloads on its GPUs, the counterpart to NVIDIA's CUDA.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.18.0 https://github.com/ollama/ollama.git
$ git checkout v0.18.0
└──▷ TRY IT
Run an automated code review in a GitHub Actions GitHub Actions GitHub's built-in CI/CD system that runs automated workflows triggered by repository events, letting cyber tools automate testing, building, and releasing without managing separate pipeline infrastructure.
copy
$ ollama launch claude \
--model glm-5:cloud \
--yes \
-- "Do a quick code review of this pull request and respond on GitHub with a comment summarizing your feedback." Use Ollama as the auth and model provider when setting up OpenClaw, pointing it at NVIDIA's new Nemotron-3-Super cloud model.
copy
$ openclaw onboard \
--auth-choice ollama \
--custom-model-id nemotron-3-super:cloud Run Nemotron-3-Super via Ollama's cloud for agentic reasoning tasks without a local 96GB VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory.
copy
$ ollama run nemotron-3-super:cloud
Adds --yes flag to ollama launch for non-interactive/scripted use (GitHub Actions, CI pipelines, etc.) Adds Nemotron-3-Super (122B) model with strong reasoning and tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. Ollama can now be selected as an authentication and model provider during OpenClaw onboarding Cloud models (:cloud tag) no longer require ollama pull — they connect automatically Ships with ROCm 7, bringing updated AMD GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
+2 more
show less
Up to 10x lower latency for MiniMax-M2.5 and up to 2x for Qwen3.5 on Ollama's cloud for North American customers Up to 2x faster speeds with Kimi-K2.5 via OpenClaw
7 details in Everything
└──▷ BREAKING ON UPGRADE
! ROCm 7 is now bundled; AMD GPU users must update drivers to the latest version for continued GPU support.
Ollama v0.17.8-rc4 adds MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. int4 int4 Four-bit integers used to represent quantized model values. Scaling maps these small integers back to approximate original values, reducing memory while losing some precision. ROCm ROCm AMD's open software stack for running AI and compute workloads on its GPUs, the counterpart to NVIDIA's CUDA.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.17.8-rc4 https://github.com/ollama/ollama.git
$ git checkout v0.17.8-rc4
Updates ROCm support on Linux to v7.2. Adds MLX int4 groupsize 64 quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. MLX runner now reads model parameters directly from the Modelfile during model creation. Removes requirement to pull stubs for cloud models, streamlining cloud model usage.
4 details in Everything
Ollama v0.17.7 adds thinking-level support in the API and context-length compaction for ollama launch.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.17.7 https://github.com/ollama/ollama.git
$ git checkout v0.17.7
└──▷ TRY IT
Control reasoning depth on a thinking model by passing a named thinking level through the API instead of a raw token budget.
copy
$ curl http://localhost:11434/api/chat -d '{"model": "qwq", "options": {"thinking": "medium"}, "messages": [{"role": "user", "content": "Explain RSA encryption."}]}'
Supports named thinking levels (e.g., "medium") for thinking models via Ollama's API. Adds context-length parameter to enable compaction when using ollama launch.
2 details in Everything
Ollama v0.17.5 adds Qwen3.5 model support and peak memory visibility in the MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.17.5 https://github.com/ollama/ollama.git
$ git checkout v0.17.5
└──▷ TRY IT
Pull and run a small Qwen3.5 model while monitoring peak MLX memory usage during inference.
copy
$ ollama pull qwen3.5:4b && ollama run --verbose qwen3.5:4b
Adds Qwen3.5 model series (0.8B, 2B, 4B, 9B parameters) to the Ollama library. ollama run --verbose now shows peak memory usage when using Ollama's MLX engine.
2 details in Everything
Ollama v0.17.4 adds Qwen 3.5 and LFM 2 models plus tool call indices in parallel tool calls.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.17.4 https://github.com/ollama/ollama.git
$ git checkout v0.17.4
└──▷ TRY IT
Pull and run the new Qwen 3.5 multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
copy
$ ollama run qwen3.5 Pull and run LFM2 for efficient on-device deployment with a large-scale hybrid model.
Adds Qwen 3.5, a multimodal open-source model family, available via ollama pull qwen3.5. Adds LFM 2, a hybrid on-device model family including LFM2-24B-A2B (24B params, efficient inference), available via ollama pull lfm2. Includes tool call indices in parallel tool call responses, enabling callers to correlate multiple simultaneous tool invocations.
3 details in Everything
└──▷ BREAKING ON UPGRADE
! Users on v0.17.1 will not receive this update automatically and must re-download Ollama manually.
Ollama v0.17.1 adds Nemotron architecture, web search for tool-capable models, and auto-update download controls.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.17.1 https://github.com/ollama/ollama.git
$ git checkout v0.17.1
Adds support for the Nemotron model architecture in Ollama's engine. Enables web search capabilities for models that support tools via the Ollama app. Adds configuration option to disable automatic update downloading. Improves LFM2 and LFM2.5 model support in Ollama's engine. MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
5 details in Everything
└──▷ BREAKING ON UPGRADE
! ollama create no longer defaults to affine quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. quantized quantized Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
Ollama v0.17.0 adds ollama launch openclaw for one-command OpenClaw setup with web search and open model support.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.17.0 https://github.com/ollama/ollama.git
$ git checkout v0.17.0
└──▷ TRY IT
Spin up a fully configured OpenClaw instance in one step to start using open models like Kimi-K2.5 without manual setup.
copy
$ ollama launch openclaw
New ollama launch openclaw command installs and configures OpenClaw automatically, enabling use with open models like Kimi-K2.5, GLM-5, and Minimax-M2.5. Adds web search capability in OpenClaw when using cloud models, allowing internet-augmented responses. macOS and Windows apps now automatically select context length context length The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory.
3 details in Everything
Ollama v0.16.3 adds Cline CLI integration, always-on model picker, and Gemma 3/Llama/Qwen 3 MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.16.3 https://github.com/ollama/ollama.git
$ git checkout v0.16.3
└──▷ TRY IT
Quickly spin up Cline with a local model — useful when you want an AI coding assistant running entirely on-device.
copy
$ ollama launch cline
New ollama launch cline subcommand launches the Cline CLI integration directly. ollama launch <integration> now always shows the model picker, making model selection explicit on every launch.Adds Gemma 3, Llama, and Qwen 3 architectures to the MLX runner, enabling local Apple Silicon acceleration for those model families.
3 details in Everything
Ollama v0.16.2 adds web search for cloud models and a new setting to block cloud data egress.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.16.2 https://github.com/ollama/ollama.git
$ git checkout v0.16.2
└──▷ TRY IT
Keep all inference on-device — block cloud model access entirely for air-gapped or sensitive environments.
copy
$ OLLAMA_NO_CLOUD=1 ollama serve Run a live web-augmented query through Claude without leaving the Ollama CLI.
copy
$ ollama launch claude:cloud
Adds web search support when running ollama launch claude with :cloud models. New OLLAMA_NO_CLOUD=1 environment variable (and app setting) to disable cloud models and keep data on-device.
2 details in Everything
Ollama v0.16.1 lets image generation models respect OLLAMA_LOAD_TIMEOUT and improves install UX on macOS and Windows.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.16.1 https://github.com/ollama/ollama.git
$ git checkout v0.16.1
Image generation models now honor the OLLAMA_LOAD_TIMEOUT environment variable, enabling timeout control for slow-loading diffusion models. macOS curl install script no longer prompts for a password unless elevation is actually required. Windows iem install script now displays progress during installation.
3 details in Everything
Ollama v0.16.0 adds a new ollama launch command, Ctrl+G editor integration, and two new frontier models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.16.0 https://github.com/ollama/ollama.git
$ git checkout v0.16.0
└──▷ TRY IT
Launch the Pi app with a local model without manual setup — useful for quickly spinning up a model-backed app.
Open your $EDITOR mid-conversation to compose or edit a long or complex prompt comfortably, then send it on save.
copy
$ ollama run llama3
# At the >>> prompt, press Ctrl+G to open the prompt in your text editor
New ollama launch command lets users start apps (e.g., Pi) pre-connected to a local model. Ctrl+G keybinding opens an external text editor for editing prompts during an interactive model session. MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. Adds GLM-5 (744B total / 40B active MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
4 details in Everything
ollama launch now auto-downloads missing models instead of erroring out
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.15.6 https://github.com/ollama/ollama.git
$ git checkout v0.15.6
Enables ollama launch to automatically download missing models on demand rather than returning an error.
1 detail in Everything
Ollama v0.15.5 adds sub-agent support, browser-based sign-in, VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.15.5 https://github.com/ollama/ollama.git
$ git checkout v0.15.5
└──▷ TRY IT
Pass extra flags to a launched tool — useful when resuming a paused coding session via Claude.
copy
$ ollama launch claude -- --resume
Adds argument passthrough to ollama launch, enabling flags like --resume to be forwarded to the launched tool. Adds sub-agent support to ollama launch for planning, deep research, and similar multi-step agentic workflows. ollama signin now opens a browser window to the connect page for easier authentication.Adds GLM-4.7-Flash support on Ollama's experimental MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. Adds new model Qwen3-Coder-Next, optimized for agentic coding workflows and local development.
+1 more
show less
Adds new model GLM-OCR OCR Optical character recognition, converting text in an image or scanned document into machine-readable characters. It lets software search or process text that was stored as pixels. multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
6 details in Everything
Ollama v0.15.4 adds automatic OpenClaw onboarding flow to ollama launch openclaw.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.15.4 https://github.com/ollama/ollama.git
$ git checkout v0.15.4
└──▷ TRY IT
Run OpenClaw for the first time and be guided through its onboarding flow automatically.
copy
$ ollama launch openclaw
Ollama v0.15.3: ollama launch now respects OLLAMA_HOST and improves Ministral tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.15.3 https://github.com/ollama/ollama.git
$ git checkout v0.15.3
└──▷ TRY IT
Launch a model against a non-default Ollama host (e.g., a remote server) without extra flags — useful in multi-host or CI environments.
copy
$ OLLAMA_HOST=https://my-ollama-server:11434 ollama launch openclaw
Enables ollama launch to use the OLLAMA_HOST environment variable when connecting, so custom host configurations are honored automatically. Improves tool calling support for Ministral models.
2 details in Everything
└──▷ BREAKING ON UPGRADE
! The ollama launch clawdbot subcommand is renamed to ollama launch openclaw; any scripts or workflows invoking ollama launch clawdbot will break.
Ollama v0.15.2 adds ollama launch clawdbot to spin up Clawdbot using local Ollama models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.15.2 https://github.com/ollama/ollama.git
$ git checkout v0.15.2
└──▷ TRY IT
Launch a Clawdbot instance backed by your local Ollama models without any external API dependency.
copy
$ ollama launch clawdbot
New ollama launch clawdbot command launches Clawdbot powered by locally running Ollama models.
1 detail in Everything
Ollama v0.15.0 adds ollama launch to connect local models with Claude Code, Codex, OpenCode, and Droid instantly.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.15.0 https://github.com/ollama/ollama.git
$ git checkout v0.15.0
New ollama launch command integrates Ollama models with Claude Code, Codex, OpenCode, and Droid without separate configuration. Supports <kbd>Ctrl</kbd>+<kbd>J</kbd> and <kbd>Shift</kbd>+<kbd>Enter</kbd> for inserting newlines in ollama run interactive sessions.
2 details in Everything
Ollama v0.14.3 adds image-generation support via /api/generate and two new text-to-image models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.14.3 https://github.com/ollama/ollama.git
$ git checkout v0.14.3
└──▷ TRY IT
Generate an image from a text prompt using the /api/generate API — no separate image endpoint needed.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "x/z-image-turbo", "prompt": "a photorealistic mountain landscape at golden hour"}'
Adds Z-Image Turbo, a 6B-parameter text-to-image model from Alibaba's Tongyi Lab for high-quality photorealistic image generation. Adds Flux.2 Klein, Black Forest Labs' fastest image-generation model to date. Adds GLM-4.7-Flash, a 30B-class model optimized for lightweight deployment balancing performance and efficiency. Adds LFM2.5-1.2B-Thinking, a hybrid model family designed for on-device deployment. Enables image generation via the existing /api/generate API endpoint.
5 details in Everything
Ollama v0.14.2 adds multiline CLI input, improved OpenResponses API conformance, and the TranslateGemma 55-language model.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.14.2 https://github.com/ollama/ollama.git
$ git checkout v0.14.2
Supports multiline input in the Ollama CLI via Shift+Enter or Ctrl+J. Improves /v1/responses API conformance to the OpenResponses specification. Adds TranslateGemma, a new collection of open translation models built on Gemma 3 supporting 55 languages.
3 details in Everything
Ollama v0.14.1 adds experimental image generation model support on macOS and Linux (CUDA CUDA NVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.14.1 https://github.com/ollama/ollama.git
$ git checkout v0.14.1
└──▷ TRY IT
Generate images interactively using the first available experimental image generation model.
copy
$ ollama run x/z-image-turbo
Ollama v0.14.0 adds an agent loop with bash tool, Anthropic API compatibility, image generation, and Modelfile version pinning.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.14.0 https://github.com/ollama/ollama.git
$ git checkout v0.14.0
└──▷ TRY IT
Run an agent loop with bash tool access — useful for tasks that require multi-step reasoning and shell execution.
copy
$ ollama run --experimental llama3.2 Call Ollama from an Anthropic-compatible client or SDK by pointing it at the /v1/messages endpoint.
copy
$ curl http://localhost:11434/v1/messages \
-H 'Content-Type: application/json' \
-d '{"model": "llama3.2", "max_tokens": 256, "messages": [{"role": "user", "content": "Hello!"}]}'
New ollama run --experimental flag opens an agent loop CLI with a built-in bash tool for agentic workflows. Adds Anthropic API compatibility via the /v1/messages endpoint, enabling Anthropic-compatible clients to target Ollama. New REQUIRES Modelfile command lets model authors declare the minimum Ollama version required to run the model. Adds experimental support for image generation models powered by MLX MLX Apple's machine-learning framework for Apple silicon, which uses the unified memory shared by CPU and GPU to run and fine-tune models on a Mac. Adds Swift source code syntax highlighting in the Ollama app.
+1 more
show less
Linux install bundles now use zst compression for faster downloads.
6 details in Everything
Ollama v0.13.5 adds BERT BERT A transformer encoder trained to understand text using context on both sides of a word. Its representations are useful for search, classification and other language-understanding tasks. function-calling function-calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.13.5 https://github.com/ollama/ollama.git
$ git checkout v0.13.5
Adds built-in renderer and tool parsing for DeepSeek-V3.1, enabling structured function-calling workflows with that model. New functiongemma model available: a Gemma 3 270M variant fine-tuned explicitly for function calling.
2 details in Everything
Ollama v0.13.4 adds Nemotron 3 Nano and OLMo 3/3.1 models, with Flash Attention Flash Attention A faster, memory-efficient way to compute attention on a GPU by working in on-chip memory tiles. Most training and serving stacks use it to handle long contexts.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.13.4 https://github.com/ollama/ollama.git
$ git checkout v0.13.4
Adds Nemotron 3 Nano, an efficient open model optimized for agentic workloads. Adds OLMo 3 and OLMo 3.1, open language models pre-trained on the Dolma 3 dataset and post-trained on Dolci datasets. Enables Flash Attention automatically by default for all models, improving inference performance.
3 details in Everything
Ollama v0.13.3 adds Devstral-Small-2, rnj-1, and nomic-embed-text-v2-moe models plus improved embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.13.3 https://github.com/ollama/ollama.git
$ git checkout v0.13.3
Adds Devstral-Small-2 (24B), a code-focused model optimized for multi-file editing and software engineering agents. Adds rnj-1, an 8B dense model from Essential AI optimized for code and STEM tasks. Adds nomic-embed-text-v2-moe, a multilingual Mixture-of-Experts Mixture-of-Experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. embedding model embedding model Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. Improves truncation logic for the /api/embed and /v1/embeddings endpoints.
4 details in Everything
Ollama v0.13.2 enables flash attention by default for vision models and adds Qwen3-Next.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.13.2 https://github.com/ollama/ollama.git
$ git checkout v0.13.2
└──▷ TRY IT
Run the new Qwen3-Next model for fast, parameter-efficient inference.
copy
$ ollama run qwen3-next
Adds Qwen3-Next model, the first in the Qwen3-Next series, optimized for parameter efficiency and inference speed. Enables flash attention by default for vision models (mistral-3, gemma3, qwen3-vl, and others), improving memory utilization and performance when processing image inputs.
2 details in Everything
Ollama v0.13.1 adds two new Mistral models, tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.13.1 https://github.com/ollama/ollama.git
$ git checkout v0.13.1
└──▷ TRY IT
Run the new edge-optimized Ministral-3 model locally for lightweight inference tasks.
copy
$ ollama run ministral-3 Use cogito-v2.1 with tool calling enabled for agentic workflows.
copy
$ ollama run cogito-v2.1
Adds Ministral-3, an edge-optimized model designed to run on a wide range of hardware. Adds Mistral-Large-3, a multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. mixture-of-experts mixture-of-experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. Enables tool calling support for the cogito-v2.1 model. Adds thinking and tool parsing for cogito-v2.1. Switches nomic-embed-text to use Ollama's native engine by default.
5 details in Everything
Ollama v0.13.0 adds DeepSeek-OCR OCR Optical character recognition, converting text in an image or scanned document into machine-readable characters. It lets software search or process text that was stored as pixels.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.13.0 https://github.com/ollama/ollama.git
$ git checkout v0.13.0
└──▷ TRY IT
Extract and convert a scanned document to markdown using DeepSeek-OCR with layout grounding.
copy
$ ollama run deepseek-ocr "/path/to/image\n<|grounding|>Convert the document to markdown." Benchmark a local model's token throughput to baseline performance or compare after hardware changes.
copy
$ go run ./cmd/bench -model gpt-oss:20b
Adds DeepSeek-OCR model support for document OCR, layout-grounded extraction, figure parsing, and markdown conversion from images. Adds Cogito-V2.1 instruction-tuned model, currently the top open-weight LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. New bench CLI tool for benchmarking local model performance (built from the Ollama repo with Go). Supports DeepSeek-V3.1 architecture natively in Ollama's engine.
4 details in Everything
Ollama v0.12.11 adds log probability support to its API and opt-in Vulkan GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.11 https://github.com/ollama/ollama.git
$ git checkout v0.12.11
└──▷ TRY IT
Retrieve per-token log probabilities to evaluate model confidence on a generation — useful for perplexity scoring or classification pipelines.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "gemma3", "prompt": "Why is the sky blue?", "logprobs": true}' Inspect the top 3 alternative tokens at each position to understand model uncertainty or build autocomplete candidates.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "gemma3", "prompt": "Why is the sky blue?", "logprobs": true, "top_logprobs": 3}' Enable Vulkan acceleration to run models on AMD, Intel, or integrated GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. CUDA CUDA NVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first.
copy
$ OLLAMA_VULKAN=1 ollama serve
Adds logprobs field to Ollama's generate API and OpenAI-compatible API, returning per-token log probabilities and raw bytes for classification, evaluation, and perplexity tasks. Adds top_logprobs parameter to expose the N most-likely alternative tokens at each position, enabling token-level introspection. Adds opt-in Vulkan GPU acceleration via OLLAMA_VULKAN=1, broadening hardware support to AMD, Intel, and integrated GPUs. Adds WebP image support in Ollama's desktop app. Ollama now prefers dedicated GPUs over iGPUs when scheduling models.
5 details in Everything
Ollama v0.12.10 adds embedding model embedding model Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. ollama run, tool call IDs in the chat API, and Vulkan GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.10 https://github.com/ollama/ollama.git
$ git checkout v0.12.10
└──▷ TRY IT
Generate a vector embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
copy
$ ollama run embeddinggemma "Hello world" Pipe text into an embedding model from a file or another command for batch or scripted embedding workflows.
copy
$ echo "Hello world" | ollama run embeddinggemma
Enables ollama run to generate vector embeddings vector embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. embedding models embedding models Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. Returns tool call IDs from the /api/chat API endpoint. Enables flash attention for Vulkan GPU backends. Adds Vulkan memory detection for Intel GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Shows login instructions when switching to a cloud model in interactive mode.
5 details in Everything
Ollama v0.12.8 adds flash attention for qwen3-vl and automatic model unload before removal.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.8 https://github.com/ollama/ollama.git
$ git checkout v0.12.8
Enables flash attention by default for qwen3-vl, improving inference performance.
1 detail in Everything
Ollama v0.12.7 adds Qwen3-VL and MiniMax-M2 models, file uploads in the app, adjustable thinking levels, and OpenAI-compatible embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.7 https://github.com/ollama/ollama.git
$ git checkout v0.12.7
└──▷ TRY IT
Request base64-encoded embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
copy
$ curl http://localhost:11434/v1/embeddings -H 'Content-Type: application/json' -d '{"model": "nomic-embed-text", "input": "Detect lateral movement in auth logs", "encoding_format": "base64"}'
Adds Qwen3-VL multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. Adds MiniMax-M2, a 230B parameter model optimized for coding and agentic workflows, available on Ollama's cloud. New app UI: attach one or multiple files when prompting a model. New app UI: adjust thinking levels for gpt-oss models to tune response depth. The OpenAI-compatible /v1/embeddings endpoint now supports the encoding_format parameter.
+2 more
show less
Extends tool-call parsing to handle responses that don't conform to the standard {"name": name, "arguments": args} format. New API documentation site launched at https://docs.ollama.com/api.
7 details in Everything
Ollama v0.12.6 adds tool-call search support, default flash attention for Gemma 3, and experimental Vulkan GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.6 https://github.com/ollama/ollama.git
$ git checkout v0.12.6
Adds web/tool-call search capability when running DeepSeek-V3.1, Qwen3, and other tool-calling tool-calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. Enables flash attention by default for Gemma 3, improving performance and memory utilization. Introduces experimental Vulkan GPU backend (build-from-source only), extending support to AMD and Intel GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
3 details in Everything
Ollama v0.12.5 adds structured output structured output Forcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. /api/chat
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.5 https://github.com/ollama/ollama.git
$ git checkout v0.12.5
└──▷ TRY IT
Use structured output with a thinking model (e.g. deepseek-r1) to get schema-constrained JSON responses from the /api/chat endpoint
copy
$ curl http://localhost:11434/api/chat -d '{"model": "deepseek-r1", "messages": [{"role": "user", "content": "Extract the name and age from: John is 30 years old."}], "format": {"type": "object", "properties": {"name": {"type": "string"}, "age": {"type": "integer"}}, "required": ["name", "age"]}, "stream": false}'
Supports structured outputs structured outputs Forcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. /api/chat API
1 detail in Everything
└──▷ BREAKING ON UPGRADE
! macOS 12 Monterey and macOS 13 Ventura are no longer supported.! AMD gfx900 and gfx906 (MI50, MI60, etc) GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. ROCm ROCm AMD's open software stack for running AI and compute workloads on its GPUs, the counterpart to NVIDIA's CUDA.
Ollama v0.12.4 enables flash attention by default for Qwen3 models and adds a new override to disable it.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.4 https://github.com/ollama/ollama.git
$ git checkout v0.12.4
└──▷ TRY IT
Disable flash attention for a Qwen3 model if you encounter compatibility issues, overriding the new default.
copy
$ OLLAMA_FLASH_ATTENTION=0 ollama run qwen3
Enables flash attention by default for Qwen3 and Qwen3 Coder models, improving inference performance without manual configuration. Adds support for overriding default flash attention via OLLAMA_FLASH_ATTENTION=0 for models where it is enabled by default.
2 details in Everything
└──▷ BREAKING ON UPGRADE
! macOS 12 Monterey and macOS 13 Ventura are no longer supported.! AMD gfx900 and gfx906 (MI50, MI60, etc.) GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. ROCm ROCm AMD's open software stack for running AI and compute workloads on its GPUs, the counterpart to NVIDIA's CUDA.
Ollama v0.12.3 adds DeepSeek-V3.1 and Kimi-K2 models plus sign-in URL prompting on push.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.3 https://github.com/ollama/ollama.git
$ git checkout v0.12.3
└──▷ TRY IT
Run DeepSeek-V3.1-Terminus on Ollama's cloud when you need a hybrid thinking/non-thinking model without local VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory.
copy
$ ollama run deepseek-v3.1:671b-cloud Run Kimi-K2's trillion-parameter MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
copy
$ ollama run kimi-k2:1t-cloud
Adds DeepSeek-V3.1-Terminus, a hybrid model supporting both thinking and non-thinking modes, runnable locally or on Ollama's cloud. Adds Kimi-K2-Instruct-0905, a 1-trillion-parameter MoE model with 32B activated parameters, available via cloud. ollama push now outputs a sign-in URL when authentication is required.
3 details in Everything
Ollama v0.12.2 adds a web search API to ground model responses in live data and expands new-engine model support.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.2 https://github.com/ollama/ollama.git
$ git checkout v0.12.2
New web search API lets models retrieve live web results to reduce hallucinations and improve accuracy, with a free tier for individuals and higher rate limits via Ollama Cloud. Qwen3 architecture models, including MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. Supports multi-regex pretokenizers in Ollama's new engine. Ollama's new engine can load tensors by matching a prefix or suffix.
4 details in Everything
Ollama v0.12.1 adds Qwen3 Embedding model Embedding model Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.1 https://github.com/ollama/ollama.git
$ git checkout v0.12.1
└──▷ TRY IT
Generate embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. Embedding Embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. RAG RAG Retrieval-augmented generation: looking up relevant documents and adding them to the prompt before the model answers. It grounds responses in private or current data without retraining. semantic search semantic search Searching by meaning rather than requiring the same words. It commonly compares vector representations to find related content, even when the wording differs.
copy
$ ollama pull qwen3-embedding
Adds Qwen3 Embedding, a state-of-the-art open embedding model from the Qwen team, available via ollama pull qwen3-embedding. Enables tool calling support for Qwen3-Coder models.
2 details in Everything
Ollama v0.12.0 adds cloud-hosted large models, native BERT BERT A transformer encoder trained to understand text using context on both sides of a word. Its representations are useful for search, classification and other language-understanding tasks.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.12.0 https://github.com/ollama/ollama.git
$ git checkout v0.12.0
└──▷ TRY IT
Run a cloud-hosted large model when local hardware is insufficient for multi-hundred-billion-parameter inference.
copy
$ ollama run qwen3-coder:480b-cloud
Adds cloud model support (preview): run massive models like qwen3-coder:480b on datacenter-grade hardware with the -cloud tag suffix. Supports BERT-architecture models natively on Ollama's own engine. Supports Qwen3-architecture models natively on Ollama's own engine. Skips parsing an initial <think> block when it is provided in the prompt for /api/generate, enabling cleaner reasoning-model workflows.
4 details in Everything
Ollama v0.11.11 adds CUDA CUDA NVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first. dimensions field for embed requests.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.11.11 https://github.com/ollama/ollama.git
$ git checkout v0.11.11
└──▷ TRY IT
Request embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. vector stores vector stores A database built to store embeddings and find the nearest ones to a query quickly. It is the retrieval half of most RAG systems. embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
copy
$ curl http://localhost:11434/api/embed -d '{"model": "nomic-embed-text", "input": "Why is the sky blue?", "dimensions": 512}'
Supports CUDA 13 for GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Adds dimensions field to embed requests, enabling control over output embedding size. Enables new memory estimates in Ollama's new engine by default, improving resource planning for hybrid and recurrent models. Adds Cmd+/- zoom controls for text sizing in Ollama's desktop app. Enables copying of assistant messages in Ollama's desktop app.
5 details in Everything
Ollama v0.11.10 adds support for EmbeddingGemma, a new best-in-class open embedding model embedding model Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.11.10 https://github.com/ollama/ollama.git
$ git checkout v0.11.10
Ollama v0.11.9 overlaps GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. CPU CPU The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.11.9 https://github.com/ollama/ollama.git
$ git checkout v0.11.9
Improves performance by overlapping GPU and CPU computations during inference.
1 detail in Everything
Ollama v0.11.8 enables flash attention by default for gpt-oss on supported hardware.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.11.8 https://github.com/ollama/ollama.git
$ git checkout v0.11.8
Ollama v0.11.7 adds DeepSeek-V3.1 support with hybrid thinking and cloud Turbo mode preview.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.11.7 https://github.com/ollama/ollama.git
$ git checkout v0.11.7
└──▷ TRY IT
Run DeepSeek-V3.1 with thinking enabled to get chain-of-thought chain-of-thought Intermediate reasoning a model writes out before its final answer. Reasoning models generate it automatically and it is often billed as extra output tokens.
copy
$ curl http://localhost:11434/api/chat -d '{"model": "deepseek-v3.1", "messages": [{"role": "user", "content": "why is the sky blue?"}], "think": true}' Run DeepSeek-V3.1 on Ollama's cloud hardware when local VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory.
copy
$ OLLAMA_HOST=ollama.com ollama run deepseek-v3.1 Toggle thinking on or off interactively during a CLI chat session with a hybrid-thinking model.
copy
$ /set think
# or to disable:
/set nothink
Adds DeepSeek-V3.1 (671B parameter model) to the Ollama model library. Supports hybrid thinking on compatible models: enable or disable chain-of-thought reasoning per request via the think field in the API or /set think//set nothink CLI commands. Introduces Turbo mode (preview): routes inference for large models like DeepSeek-V3.1 to cloud hardware via OLLAMA_HOST=ollama.com.
3 details in Everything
Ollama v0.11.5 adds opt-in improved VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory. GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. CPU CPU The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.11.5 https://github.com/ollama/ollama.git
$ git checkout v0.11.5
└──▷ TRY IT
Try the new VRAM estimator before it becomes the default — useful if you're hitting out-of-memory out-of-memory An out-of-memory condition: a program requests more memory than the system or device can provide. The allocation fails or the operating system may stop the process.
copy
$ OLLAMA_NEW_ESTIMATES=1 ollama serve Speed up inference on CPU-only hosts (e.g., a server with no GPU) by enabling flash attention, which was previously a GPU-only optimization.
copy
$ OLLAMA_FLASH_ATTENTION=1 ollama serve
New opt-in memory management (OLLAMA_NEW_ESTIMATES=1) improves VRAM utilization and reduces out-of-memory errors when scheduling models on GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Improved multi-GPU scheduling reduces VRAM allocation when running across more than 2 GPUs. Enables flash attention for pure-CPU models via OLLAMA_FLASH_ATTENTION=1. The Ollama app now persists default model, Turbo, and Web Search selections across restarts.
4 details in Everything
Ollama v0.11.4 improves OpenAI-compatible API support: simultaneous content+tool calls, tool name propagation, and always-on reasoning.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.11.4 https://github.com/ollama/ollama.git
$ git checkout v0.11.4
Supports content and tool calls coexisting in the same OpenAI-compatible message, enabling richer agentic responses. Propagates the tool name when converting role=tool messages in the OpenAI-compatible layer. Always provides reasoning output in OpenAI-compatible responses.
3 details in Everything
Ollama v0.11 adds native support for OpenAI's gpt-oss 20B and 120B open-weight models with built-in web search and MXFP4 MXFP4 A four-bit floating-point format with a shared scale for each small group of values. The scales help preserve useful precision while compressing model weights and calculations. quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.11.0 https://github.com/ollama/ollama.git
$ git checkout v0.11.0
└──▷ TRY IT
Run the smaller gpt-oss model locally on a 16 GB system for reasoning or agentic tasks.
copy
$ ollama run gpt-oss:20b Run the larger gpt-oss model on an 80 GB GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
copy
$ ollama run gpt-oss:120b
Adds OpenAI gpt-oss 20B and 120B open-weight models, runnable locally via ollama run gpt-oss:20b and ollama run gpt-oss:120b. Supports native MXFP4 quantization format for gpt-oss MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. RAM RAM Working memory that holds programs and data while a computer is running. Its contents usually disappear when power is removed. Enables built-in optional web search to augment gpt-oss models with real-time information. Supports configurable reasoning effort (low, medium, high) for gpt-oss models to balance quality and latency. Exposes full chain-of-thought chain-of-thought Intermediate reasoning a model writes out before its final answer. Reasoning models generate it automatically and it is often billed as extra output tokens.
+2 more
show less
Supports function calling function calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. structured outputs structured outputs Forcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. Adds SWA (sliding window attention sliding window attention Attention that reads only a nearby window of tokens instead of the entire sequence. It reduces computation and memory for long inputs, but limits each layer's direct view of distant tokens. KV-cache KV-cache Memory where a model keeps the attention keys and values of tokens it has already processed, so each new token does not recompute them. Its size grows with context length and often limits serving.
7 details in Everything
Ollama v0.10.0 adds context-length visibility in ollama ps, WebP image support, and 10-30% multi-GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.10.0 https://github.com/ollama/ollama.git
$ git checkout v0.10.0
└──▷ TRY IT
Inspect context window context window The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it.
Adds context length context length The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. ollama ps output for loaded models. Supports WebP images in the OpenAI-compatible API. Delivers 10-30% performance improvement when using multiple GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Improves performance of gemma3n models by 2-3x. Launches redesigned desktop app for macOS and Windows.
5 details in Everything
└──▷ BREAKING ON UPGRADE
! Parallel request processing now defaults to 1 (previously higher), which will reduce throughput for workloads relying on the previous concurrent-request default.
Ollama v0.9.6 lets tool-role messages carry a tool_name field in /api/chat.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.9.6 https://github.com/ollama/ollama.git
$ git checkout v0.9.6
└──▷ TRY IT
Attribute a tool result to a specific tool by name when replying to a model's tool call in a multi-turn chat session.
copy
$ curl http://localhost:11434/api/chat -d '{"model": "llama3", "messages": [{"role": "user", "content": "What is the weather in Paris?"}, {"role": "assistant", "tool_calls": [{"function": {"name": "get_weather", "arguments": {"city": "Paris"}}}]}, {"role": "tool", "tool_name": "get_weather", "content": "Sunny, 22°C"}]}'
Supports tool_name field in messages with "role": "tool" via the /api/chat endpoint, enabling precise tool-call attribution in multi-turn agentic conversations.
1 detail in Everything
Ollama v0.9.5 adds network exposure, configurable model directory, and a faster native macOS app.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.9.5 https://github.com/ollama/ollama.git
$ git checkout v0.9.5
Enables exposing Ollama over the network so other devices (including low-powered ones) can access a centrally running instance. Supports configuring the model storage directory, enabling models to live on external drives or custom paths. Ships a native macOS application with smaller installation footprint and faster startup time. Adds NativeMind to the Community Integrations list.
4 details in Everything
└──▷ BREAKING ON UPGRADE
! Ollama for macOS now requires version 12 (Monterey) or newer; installations on older macOS versions will no longer be supported.
Ollama v0.9.4 adds network exposure and configurable model directory via the desktop apps.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.9.4 https://github.com/ollama/ollama.git
$ git checkout v0.9.4
Enables exposing the Ollama server on the network so other devices (or remote users) can access models running on a powerful local machine. Adds configurable model storage directory, allowing models to be stored on external drives or any non-default path. Delivers a native macOS app with significantly smaller install footprint and faster startup.
3 details in Everything
└──▷ BREAKING ON UPGRADE
! Ollama for macOS now requires version 12 (Monterey) or newer; installations on older macOS versions will no longer be supported.
Ollama v0.9.3 adds support for Google's Gemma 3n models, optimized for laptops, tablets, and phones.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.9.3 https://github.com/ollama/ollama.git
$ git checkout v0.9.3
└──▷ TRY IT
Run the Effective 2B Gemma 3n model locally for lightweight, multilingual inference on everyday hardware.
copy
$ ollama run gemma3n:e2b Run the Effective 4B Gemma 3n model for higher-quality multilingual responses while remaining efficient on consumer devices.
copy
$ ollama run gemma3n:e4b
Supports Gemma 3n models (e2b and e4b variants), designed for efficient on-device execution across laptops, tablets, and phones with training data covering 140+ spoken languages.
1 detail in Everything
Ollama v0.9.1 adds tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.9.1 https://github.com/ollama/ollama.git
$ git checkout v0.9.1
└──▷ TRY IT
Pull DeepSeek-R1-2508 (671B) or Magistral to start using the newly supported tool calling capability.
copy
$ ollama pull deepseek-r1:671b
Adds tool calling support for DeepSeek-R1-2508 (671B) and Magistral models. Supports disabling thinking mode on Magistral (with recommended system prompt change). New preview macOS and Windows desktop apps with network exposure, local browser access, and configurable model directory. macOS app rebuilt as a native application for smaller footprint and faster startup. Enables exposing Ollama on the network so other devices (or remote users) can reach a central Ollama host.
+2 more
show less
Allows local browser access so web applications can directly call the local Ollama API. Model storage directory is now configurable, enabling use of external drives or custom paths.
7 details in Everything
Ollama v0.9.0 adds controllable thinking mode for reasoning models, exposing chain-of-thought chain-of-thought Intermediate reasoning a model writes out before its final answer. Reasoning models generate it automatically and it is often billed as extra output tokens.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.9.0 https://github.com/ollama/ollama.git
$ git checkout v0.9.0
└──▷ TRY IT
Toggle thinking off mid-session in the CLI when you want faster responses without reasoning output.
copy
$ ollama run deepseek-r1
>>> /set nothink
Adds thinking mode support: enable or disable chain-of-thought reasoning for compatible models (DeepSeek R1, Qwen 3) independently per request. New thinking field in API chat responses separates the model's reasoning trace from its final content for easy parsing. New /set think and /set nothink CLI commands toggle thinking mode interactively during a session. New "think": true/false API parameter in /api/chat allows per-request control of thinking behavior. Adds support for DeepSeek-R1-0528, the updated 8B distilled and 671B full models with improved reasoning.
5 details in Everything
Ollama v0.8.0 adds streaming support for tool call responses.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.8.0 https://github.com/ollama/ollama.git
$ git checkout v0.8.0
Streams responses that include tool calls, enabling real-time output during tool-augmented LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Adds improved memory estimate debug information in logs when running models in Ollama's engine.
2 details in Everything
Ollama v0.7.1 adds Qwen 3 & Qwen 2 multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. ollama show.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.7.1 https://github.com/ollama/ollama.git
$ git checkout v0.7.1
Supports Alibaba's Qwen 3 and Qwen 2 architectures in Ollama's multimodal engine. ollama show now displays ... when output data is truncated.
2 details in Everything
Ollama v0.7.0 adds multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.7.0 https://github.com/ollama/ollama.git
$ git checkout v0.7.0
Supports multimodal vision models via Ollama's new engine, including Meta Llama 4, Google Gemma 3, Qwen 2.5 VL, and Mistral Small 3.1. Accepts WebP images as input to multimodal models. API now returns HTTP 405 (Method Not Allowed) instead of 404 for disallowed methods, enabling more accurate client-side error handling. Improved performance of importing safetensors safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face. ollama create. Improved prompt processing prompt processing The first phase of inference, where the model reads the whole prompt in parallel before producing its first token. It drives time-to-first-token; decode is the token-by-token phase after it. MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
5 details in Everything
Ollama v0.6.8 delivers major performance gains for Qwen3 MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.8 https://github.com/ollama/ollama.git
$ git checkout v0.6.8
Improves inference performance for Qwen3 MoE models (30b-a3b and 235b-a22b) on NVIDIA and AMD GPUs.
1 detail in Everything
Ollama v0.6.7 adds Llama 4 multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. context window context window The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.7 https://github.com/ollama/ollama.git
$ git checkout v0.6.7
└──▷ TRY IT
Pull and run Phi 4 Reasoning for complex, multi-step reasoning tasks without extra configuration.
copy
$ ollama run phi4-reasoning Pull Qwen3 MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. mixture-of-experts mixture-of-experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
Adds support for Meta's Llama 4 multimodal models, enabling image-and-text inference locally. Adds support for Microsoft's Phi 4 Reasoning and Phi 4 Mini Reasoning models for state-of-the-art chain-of-thought chain-of-thought Intermediate reasoning a model writes out before its final answer. Reasoning models generate it automatically and it is often billed as extra output tokens. Adds Qwen3 family (dense and MoE variants) to the model library. Increases default context window from its previous limit to 4096 tokens, unlocking longer conversations and documents out of the box.
4 details in Everything
Ollama v0.6.6 adds IBM Granite 3.3 and DeepCoder models, experimental faster downloader, and expanded tool-call type support.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.6 https://github.com/ollama/ollama.git
$ git checkout v0.6.6
└──▷ TRY IT
Enable the experimental faster downloader to get improved pull speeds and reliability before it becomes default.
copy
$ OLLAMA_EXPERIMENT=client2 ollama serve
Adds IBM Granite 3.3 (2B and 8B) models with 128K context length context length The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. Adds DeepCoder 14B (and 1.5B) fully open-source coder model at O3-mini level. New experimental faster model downloader with improved performance and reliability, enabled via OLLAMA_EXPERIMENT=client2. Improves performance of ollama create when importing models from Safetensors Safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face. Supports tool function parameters with either a single type or an array of types.
+2 more
show less
Includes items and $defs fields in the API to properly handle array types. Adds OpenAI-Beta headers to the CORS CORS A browser mechanism that lets a server declare which foreign origins may read its HTTP responses, enforced by the browser to block unauthorized cross-site requests from JavaScript.
7 details in Everything
Ollama v0.6.5 adds Mistral Small 3.1, a top-performing vision model in its weight class.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.5 https://github.com/ollama/ollama.git
$ git checkout v0.6.5
└──▷ TRY IT
Pull and run Mistral Small 3.1 to get vision-capable inference locally.
copy
$ ollama run mistral-small3.1
Ollama v0.6.4 adds model capability metadata to /api/show and AMD RDNA4 GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.4 https://github.com/ollama/ollama.git
$ git checkout v0.6.4
└──▷ TRY IT
Programmatically detect whether a locally pulled model supports vision before routing multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
copy
$ curl http://localhost:11434/api/show -d '{"name": "llava"}'
Adds model capability metadata (e.g. vision) to /api/show API responses. Adds AMD RDNA4 GPU support on Linux.
2 details in Everything
Ollama v0.6.3 adds sliding window attention sliding window attention Attention that reads only a nearby window of tokens instead of the entire sequence. It reduces computation and memory for long inputs, but limits each layer's direct view of distant tokens. ollama create for safetensors safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.3 https://github.com/ollama/ollama.git
$ git checkout v0.6.3
Adds sliding window attention optimizations for Gemma 3, improving inference speed and memory efficiency for long context windows context windows The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. ollama create now reports the name of unsupported architectures instead of failing silently.
2 details in Everything
Ollama v0.6.2 adds multi-image support for Gemma 3 and AMD Strix Halo GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.2 https://github.com/ollama/ollama.git
$ git checkout v0.6.2
Supports multiple images in a single Gemma 3 prompt, enabling multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. Adds support for AMD Strix Halo GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. ollama create --quantize now works when converting Gemma 3 models from safetensors safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face.
3 details in Everything
Ollama v0.6.1 adds Command A model support, verbose model inspection, and new CLI hotkeys.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.1 https://github.com/ollama/ollama.git
$ git checkout v0.6.1
└──▷ TRY IT
Inspect full model metadata and configuration details — useful when debugging model behavior or confirming quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
copy
$ ollama show --verbose llama3 Pull and run the new Command A enterprise model for high-quality reasoning tasks.
copy
$ ollama run command-a
Adds Command A, a 111B-parameter enterprise-grade model, to the Ollama library. New ollama show --verbose / ollama show -v flag prints additional model data beyond the default output. Adds Ctrl+P and Ctrl+N hotkeys for navigating history in ollama run interactive sessions.
3 details in Everything
Ollama v0.6.0 adds support for Google Gemma 3 in 1B, 4B, 12B, and 27B parameter sizes.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.6.0 https://github.com/ollama/ollama.git
$ git checkout v0.6.0
└──▷ TRY IT
Run the Gemma 3 27B model locally for a large-context reasoning task.
copy
$ ollama run gemma3:27b
Ollama v0.5.13 adds NVIDIA Blackwell support, a new default context length context length The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. function calling function calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.13 https://github.com/ollama/ollama.git
$ git checkout v0.5.13
└──▷ TRY IT
Set a server-wide default context length of 8K so all models use it without per-request configuration.
copy
$ OLLAMA_CONTEXT_LENGTH=8192 ollama serve
New OLLAMA_CONTEXT_LENGTH environment variable lets operators set the default context length server-wide without per-request overrides. Ollama is now compiled for NVIDIA Blackwell GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Adds Phi-4-Mini model with multilingual support, reasoning, mathematics, and function calling. Adds Granite-3.2-Vision, a compact vision-language model vision-language model A vision-language model that works with images and text together, such as answering questions about a picture or describing what it contains. Adds Command R7B Arabic, a model optimized for advanced Arabic language tasks targeting MENA enterprises.
+1 more
show less
Accepts requests from Visual Studio Code and Cursor via origins beginning with vscode-file://, enabling IDE-native integration.
6 details in Everything
└──▷ BREAKING ON UPGRADE
! Ubuntu 20.04, Debian 10, and RHEL RHEL Red Hat Enterprise Linux, a commercially supported Linux distribution maintained by Red Hat, chosen by cyber tools for its long support lifecycle, certified hardware compatibility, and enterprise deployment prevalence.
Ollama v0.5.12 adds Perplexity R1 1776, OpenAI-compatible tool_calls responses, and X-Stainless-Timeout header support.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.12 https://github.com/ollama/ollama.git
$ git checkout v0.5.12
└──▷ TRY IT
Pull and run the new Perplexity R1 1776 model for sensitive-topic research queries it previously refused.
copy
$ ollama run r1-1776
Adds Perplexity R1 1776 model: a post-trained DeepSeek-R1 variant with fewer topic refusals. OpenAI-compatible API now returns tool_calls in responses when a model invokes a tool. Accepts X-Stainless-Timeout as a valid header on OpenAI API endpoints.
3 details in Everything
Ollama v0.5.9 adds DeepScaleR and OpenThinker reasoning models to the library.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.9 https://github.com/ollama/ollama.git
$ git checkout v0.5.9
└──▷ TRY IT
Run the compact DeepScaleR math-reasoning model locally for step-by-step problem solving without needing a large GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
copy
$ ollama run deepscaler Pull and run OpenThinker for open-source chain-of-thought chain-of-thought Intermediate reasoning a model writes out before its final answer. Reasoning models generate it automatically and it is often billed as extra output tokens.
copy
$ ollama run openthinker
Adds DeepScaleR, a 1.5B-parameter math-reasoning model fine-tuned on Deepseek-R1-Distilled-Qwen-1.5B, available via ollama pull deepscaler. Adds OpenThinker, a fully open-source family of reasoning models distilled from DeepSeek-R1, available via ollama pull openthinker.
2 details in Everything
Ollama v0.5.8 adds AVX-512 CPU CPU The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware. GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.8 https://github.com/ollama/ollama.git
$ git checkout v0.5.8
Adds AVX-512 instruction support for additional CPU inference acceleration on compatible hardware. Enables NVIDIA and AMD GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. CPUs CPUs The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware. Adds AVX2 instruction usage when offloading to NVIDIA and AMD GPUs for improved hybrid performance. New ollama-darwin.tgz archive package replaces the previous ollama-darwin standalone binary for macOS distribution.
4 details in Everything
└──▷ BREAKING ON UPGRADE
! The ollama-darwin standalone binary is replaced by the ollama-darwin.tgz package; scripts or pipelines that download the old binary path will break.! The steps to build Ollama with GPU acceleration from source have changed; existing build procedures must be updated per the development documentation.
Ollama v0.5.7 adds native import support for Command R and Command R+ safetensor models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.7 https://github.com/ollama/ollama.git
$ git checkout v0.5.7
Supports importing Command R and Command R+ architectures directly from safetensors safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face.
1 detail in Everything
Ollama v0.5.5 adds 8 new models including DeepSeek-V3 and Phi-4, plus a faster /api/create endpoint that now accepts JSON.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.5 https://github.com/ollama/ollama.git
$ git checkout v0.5.5
Adds Phi-4 (14B, Microsoft) to the model library. Adds Command R7B (Cohere) for efficient inference on commodity GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Adds DeepSeek-V3, a 671B MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. Adds OLMo 2 (7B and 13B) trained on up to 5T tokens. Adds Dolphin 3, a general-purpose instruct-tuned model supporting coding, math, agentic, and function-calling function-calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
+4 more
show less
Adds SmallThinker, a small reasoning model fine-tuned from Qwen 2.5 3B Instruct. Adds Granite 3.1 Dense (2B and 8B) from IBM, trained on over 12 trillion tokens. Adds Granite 3.1 MoE (1B and 3B) from IBM, designed for low-latency long-context inference. Updates the /api/create endpoint to improve model conversion speed and accept a JSON object body.
9 details in Everything
└──▷ BREAKING ON UPGRADE
! The /api/create API endpoint (used by ollama create) is not backwards compatible: both the Ollama server and the ollama CLI must be version 0.5.5 or later, and the Python/JavaScript ollama library must be updated to the latest version when calling ollama.create.
Ollama v0.5.4 adds the Falcon3 family of sub-10B models optimised for science, math, and coding.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.4 https://github.com/ollama/ollama.git
$ git checkout v0.5.4
└──▷ TRY IT
Pull and run a Falcon3 model locally for coding or math tasks.
copy
$ ollama run falcon3
Ollama v0.5.2 adds EXAONE 3.5 bilingual models (2.4B–32B) and case-insensitive OLLAMA_ORIGINS host matching.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.2 https://github.com/ollama/ollama.git
$ git checkout v0.5.2
Adds EXAONE 3.5 instruction-tuned bilingual (English/Korean) models ranging from 2.4B to 32B parameters, developed by LG AI Research. OLLAMA_ORIGINS now matches hosts case-insensitively, simplifying origin allowlist configuration.
2 details in Everything
└──▷ BREAKING ON UPGRADE
! The Linux ollama-linux-amd64.tgz directory structure has changed — manual Linux installations must retain the new directory layout and contents of the tar file after upgrading.
Ollama v0.5.0 adds JSON-schema structured outputs structured outputs Forcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. KV cache KV cache Memory where a model keeps the attention keys and values of tokens it has already processed, so each new token does not recompute them. Its size grows with context length and often limits serving. quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.5.0 https://github.com/ollama/ollama.git
$ git checkout v0.5.0
└──▷ TRY IT
Enforce a strict JSON schema on model output so downstream code can parse it reliably without extra validation.
copy
$ curl -X POST http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{"model": "llama3.1", "messages": [{"role": "user", "content": "Tell me about Canada."}], "stream": false, "format": {"type": "object", "properties": {"name": {"type": "string"}, "capital": {"type": "string"}, "languages": {"type": "array", "items": {"type": "string"}}}, "required": ["name", "capital", "languages"]}}' Run Ollama with 4-bit KV cache quantization to fit longer context windows context windows The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
copy
$ OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q4_0 ollama serve
Supports structured outputs via JSON schema in the format parameter across the REST API REST API A web API design style where clients interact with resources over HTTP using standard methods like GET, POST, and DELETE, giving cyber tools a predictable, language-agnostic way to integrate with external services. Adds Llama 3.3 (70B) model, offering performance comparable to Llama 3.1 405B. Adds Snowflake Arctic Embed 2, a multilingual frontier embedding model embedding model Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. Experimental OLLAMA_KV_CACHE_TYPE env var enables KV cache quantization at 4-bit (q4_0), 8-bit (q8_0), or 16-bit (f16), reducing VRAM requirements for longer context windows.
4 details in Everything
Ollama v0.4.6 adds QwQ reasoning model, streaming tool calls, and SVG input validation.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.4.6 https://github.com/ollama/ollama.git
$ git checkout v0.4.6
└──▷ TRY IT
Run the new QwQ reasoning model interactively to leverage its advanced chain-of-thought chain-of-thought Intermediate reasoning a model writes out before its final answer. Reasoning models generate it automatically and it is often billed as extra output tokens.
Adds QwQ, an experimental AI reasoning model from the Qwen team, available via ollama pull qwq. Tool calls are now included in streaming responses, enabling real-time agentic workflows without waiting for full completion. Ollama now returns an explicit error when an SVG image is submitted, surfacing unsupported input early.
3 details in Everything
Ollama v0.4.5 lets ollama push print the uploaded model URL and accepts X-Stainless-Retry-Count from OpenAI clients.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.4.5 https://github.com/ollama/ollama.git
$ git checkout v0.4.5
Accepts X-Stainless-Retry-Count header used by many OpenAI API clients, improving compatibility. ollama push now prints the uploaded model URL on ollama.com after a successful push.
2 details in Everything
Ollama v0.4.4 adds Marco-o1 reasoning model and clean file-redirect output.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.4.4 https://github.com/ollama/ollama.git
$ git checkout v0.4.4
└──▷ TRY IT
Run the new Marco-o1 reasoning model for complex problem-solving tasks.
copy
$ ollama run marco-o1 Capture clean model output to a file without progress bar noise polluting the log.
copy
$ ollama run marco-o1 'Explain the steps to perform a SQL injection attack and how to defend against it' > output.txt
Adds Marco-o1, Alibaba AIDC-AI's open large reasoning model, to the Ollama library. Redirecting CLI output to a file now suppresses progress bars and spinners, enabling clean log capture.
2 details in Everything
Ollama v0.4.3 adds Tülu 3 and an updated Mistral Large with improved long context and function calling function calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.4.3 https://github.com/ollama/ollama.git
$ git checkout v0.4.3
Adds Tülu 3, a fully open-source instruction-following model family from the Allen Institute for AI, available via ollama pull tulu3. Adds updated Mistral Large with improved long context, function calling, and system prompt support, available via ollama pull mistral-large.
2 details in Everything
Ollama v0.4.2 adds NVIDIA Jetson support, three new models, and improved Modelfile error reporting.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.4.2 https://github.com/ollama/ollama.git
$ git checkout v0.4.2
└──▷ TRY IT
Install Ollama on an NVIDIA Jetson device using the standard one-liner.
copy
$ curl -fsSL https://ollama.com/install.sh | sh
Supports NVIDIA Jetson devices via the standard install script. Adds Qwen 2.5 Coder in six sizes (0.5B–32B) for code generation tasks. Adds OpenCoder (1.5B and 8B) with English and Chinese chat support. Adds Athene V2 (72B) optimised for code completion, mathematics, and log extraction. ollama create now reports line numbers for Modelfile errors, speeding up authoring.
5 details in Everything
Ollama v0.4 adds Llama 3.2 Vision (multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. Safetensors Safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.4.0 https://github.com/ollama/ollama.git
$ git checkout v0.4.0
└──▷ TRY IT
Run the 90B Llama 3.2 Vision model locally for large-scale image-understanding tasks.
copy
$ ollama run llama3.2-vision:90b Import a Safetensors model checkpoint directly without authoring a Modelfile first.
copy
$ ollama create my-model
Supports the Llama 3.2 Vision (Mllama) multimodal architecture, enabling image-understanding workloads via ollama run llama3.2-vision. Enables importing models directly from Safetensors files without requiring a Modelfile when running ollama create. Follow-on requests to vision models are now significantly faster, enabling more responsive multi-turn visual conversations.
3 details in Everything
Ollama v0.3.14 adds IBM Granite 3 MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. RAG RAG Retrieval-augmented generation: looking up relevant documents and adding them to the prompt before the model answers. It grounds responses in private or current data without retraining.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.14 https://github.com/ollama/ollama.git
$ git checkout v0.3.14
└──▷ TRY IT
Pull and run the IBM Granite 3 MoE 3B model for low-latency inference tasks.
copy
$ ollama run granite3-moe Pull and run the IBM Granite 3 Dense 8B model for tool-use or RAG-backed workflows.
copy
$ ollama run granite3-dense
Adds Granite 3 MoE (1B and 3B): IBM's first mixture-of-experts mixture-of-experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. Adds Granite 3 Dense (2B and 8B): IBM Granite models designed for tool-use, RAG, code generation, and translation. Supports the IBM Granite and GraniteMoE model architectures natively.
3 details in Everything
Ollama v0.3.13 adds Llama Guard 3 and ShieldGemma safety classification models for LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.13 https://github.com/ollama/ollama.git
$ git checkout v0.3.13
Adds Llama Guard 3 support: Meta's fine-tuned content safety classification models for screening LLM inputs and outputs. Adds ShieldGemma support: Google DeepMind's instruction-tuned models for evaluating text prompts and responses against defined safety policies.
2 details in Everything
Ollama v0.3.12 adds Llama 3.2 and Qwen 2.5 Coder support, plus ARM ARM A family of RISC processor architectures designed by Arm Holdings and licensed to chip makers worldwide; tools target it to run natively on mobile, embedded, and increasingly server and desktop hardware.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.12 https://github.com/ollama/ollama.git
$ git checkout v0.3.12
└──▷ TRY IT
Run Meta's new lightweight Llama 3.2 3B model locally for fast, low-resource inference.
copy
$ ollama run llama3.2 Pull and run Qwen 2.5 Coder for code generation and reasoning tasks.
copy
$ ollama run qwen2.5-coder
Adds Llama 3.2 (1B and 3B) models from Meta, optimized for small-footprint deployments. Adds Qwen 2.5 Coder, a code-specialized model series with improved code generation and reasoning. Supports ARM Windows machines.
3 details in Everything
Ollama v0.3.11 adds ollama stop to unload running models and ships five new models including Qwen 2.5 and Bespoke-Minicheck.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.11 https://github.com/ollama/ollama.git
$ git checkout v0.3.11
└──▷ TRY IT
Free GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. CPU CPU The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware.
copy
$ ollama stop mistral-small
New ollama stop command to unload a running model from memory on demand. Adds Solar-Pro-Preview (22B), Qwen 2.5 (up to 128K context), Bespoke-Minicheck (fact-checking), Mistral-Small (22B), and Reader-LM (HTML-to-Markdown) to the model library. Shows an error when importing a model with an invalid number of tokens in the vocabulary, catching bad imports early. The ollama/ollama container image now starts almost immediately, cutting startup time by ~5 seconds.
4 details in Everything
Ollama v0.3.10 adds MiniCPM-V multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. CUDA CUDA NVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.10 https://github.com/ollama/ollama.git
$ git checkout v0.3.10
Adds Yi-Coder, an open-source code language model series delivering state-of-the-art coding performance under 10B parameters. Adds DeepSeek-V2.5, combining DeepSeek-V2 general chat and DeepSeek-Coder-V2 coding capabilities in a single model. Automatically falls back to CUDA 11 for older NVIDIA drivers incompatible with CUDA 12, expanding GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
3 details in Everything
└──▷ BREAKING ON UPGRADE
! The OpenAI-compatible chat and completions APIs no longer scale temperature and frequency_penalty, which will change generation behavior for any integration that relied on Ollama's previous scaling of these parameters.
Ollama v0.3.9 adds path support to OLLAMA_HOST and improves error messaging for unsupported models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.9 https://github.com/ollama/ollama.git
$ git checkout v0.3.9
└──▷ TRY IT
Route Ollama through a reverse proxy that serves it at a subpath rather than the root.
copy
$ OLLAMA_HOST=https://myproxy.example.com/ollama ollama run llama3.1
Supports URL paths in OLLAMA_HOST, enabling deployments behind reverse proxies with subpath routing. Shows improved error messages when attempting to run unsupported models.
2 details in Everything
Ollama v0.3.7 adds Hermes 3, Phi 3.5, and SmolLM models plus CUDA CUDA NVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first. GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.7 https://github.com/ollama/ollama.git
$ git checkout v0.3.7
└──▷ TRY IT
Run the new Hermes 3 model with tool-calling tool-calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
copy
$ ollama run hermes3
Adds Hermes 3 (Nous Research flagship LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. ollama run hermes3. Adds Phi 3.5, a 3.8B-parameter lightweight model competitive with larger models. Adds SmolLM family of small models at 135M, 360M, and 1.7B parameters trained on a new high-quality dataset. Adds CUDA 12 support, improving performance by up to 10% on newer NVIDIA GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Improves throughput of ollama pull and ollama push on slower network connections.
+1 more
show less
Linux distribution now ships as a tar.gz archive bundling the ollama binary with required libraries.
6 details in Everything
└──▷ BREAKING ON UPGRADE
! Ollama on Linux is now distributed as a tar.gz file instead of a standalone binary; existing install scripts or package pipelines that expected a bare binary will need to be updated to extract the archive.
Ollama v0.3.6 adds Phi-3 Safetensors Safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face. GGUF GGUF The single-file model format used by llama.cpp and Ollama, holding the weights, usually quantized, together with the tokenizer and settings needed to run the model locally.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.6 https://github.com/ollama/ollama.git
$ git checkout v0.3.6
Supports importing Phi-3 models from Safetensors via ollama create. Shows progress information during GGUF file imports in ollama create. Faster GGUF file importing by minimizing file copies.
3 details in Everything
Ollama v0.3.4 adds a batch embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. embedding models embedding models Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.4 https://github.com/ollama/ollama.git
$ git checkout v0.3.4
└──▷ TRY IT
Generate normalized embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. vector store vector store A database built to store embeddings and find the nearest ones to a query quickly. It is the retrieval half of most RAG systems.
copy
$ curl http://localhost:11434/api/embed -d '{"model": "bge-m3", "input": ["Why is the sky blue?", "Why is the grass green?"]}' Enforce strict input-length limits when embedding — the call errors rather than silently truncating, preventing silent data loss in pipelines.
copy
$ curl http://localhost:11434/api/embed -d '{"model": "bge-large", "input": ["Some long document..."], "truncate": false}'
New /api/embed endpoint supports batch embedding generation — submit multiple documents in a single request. Normalized embeddings now returned by default from /api/embed, improving similarity search accuracy. New truncate parameter on /api/embed lets callers enforce strict input-length control. Embedding API responses now include load_duration, total_duration, and prompt_eval_count performance metrics. Adds BGE-M3, a multilingual multi-granularity large embedding model embedding model Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
+3 more
show less
Adds BGE-Large, a large English embedding model. Adds Paraphrase-Multilingual, a multilingual embedding model supporting 50+ languages. NUMA support is now autodetected to improve inference performance on multi-socket systems.
8 details in Everything
Ollama v0.3.3 adds timing and token-count metrics to both native and OpenAI-compatible embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.3 https://github.com/ollama/ollama.git
$ git checkout v0.3.3
└──▷ TRY IT
Benchmark embedding latency and token throughput for a model without external tooling — metrics are now returned inline.
copy
$ curl http://localhost:11434/api/embed -d '{"model": "nomic-embed-text", "input": "Why is the sky blue?"}' | jq '{total_duration, load_duration, prompt_eval_count}' Track token usage from the OpenAI-compatible embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
copy
$ curl http://localhost:11434/v1/embeddings -H 'Content-Type: application/json' -d '{"model": "nomic-embed-text", "input": "Why is the sky blue?"}' | jq '.usage'
Adds total_duration, load_duration, and prompt_eval_count statistics to the /api/embed response. Adds usage metrics to the /v1/embeddings OpenAI compatibility endpoint.
2 details in Everything
Ollama v0.3.1 adds Gemma 2 2B, min_p sampling, auto-detected stop params, and tool_calls finish_reason in the OpenAI-compatible API.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.1 https://github.com/ollama/ollama.git
$ git checkout v0.3.1
└──▷ TRY IT
Run the new lightweight Gemma 2 2B model for fast, low-resource inference on-device.
copy
$ ollama run gemma2:2b
Adds Gemma 2 2B (2-billion-parameter model by Google DeepMind) to the model library via ollama run gemma2:2b Supports min_p sampling option for more controlled text generation OpenAI-compatible API now returns finish_reason: tool_calls when a tool call occurs
3 details in Everything
Ollama v0.3.0 adds tool/function-calling function-calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.3.0 https://github.com/ollama/ollama.git
$ git checkout v0.3.0
└──▷ USE IT
Let a model decide which tool to call to answer a user query — useful for building agents that dispatch to real APIs based on natural language.
python
copy
import ollama
response = ollama.chat(
model='llama3.1',
messages=[{'role': 'user', 'content': 'What is the weather in Toronto?'}],
tools=[{
'type': 'function',
'function': {
'name': 'get_current_weather',
'description': 'Get the current weather for a city',
'parameters': {
'type': 'object',
'properties': {
'city': {
'type': 'string',
'description': 'The name of the city',
},
},
'required': ['city'],
},
},
}],
)
print(response['message']['tool_calls'])
Adds tool calling tool calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. Supports Llama 3.1 (8B, 70B, 405B) with native tool-calling capability. Supports Mistral Large 2 (123B) with 128k context, code generation, tool calling, math, and multilingual reasoning. Supports Firefunction v2, an open-weights function-calling model competitive with GPT-4o function calling. Supports Llama-3-Groq-Tool-Use, a series of models optimized for tool use tool use A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions.
5 details in Everything
Ollama v0.2.8 adds Mistral Nemo (12B, 128k context) and NuExtract (3.8B info-extraction) models, plus template validation on create.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.2.8 https://github.com/ollama/ollama.git
$ git checkout v0.2.8
└──▷ TRY IT
Run the new Mistral Nemo model to leverage its 128k context window context window The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it.
copy
$ ollama run mistral-nemo Use NuExtract to pull structured data from unstructured text as part of a threat-intel or log-parsing workflow.
copy
$ ollama run nuextract
Adds Mistral Nemo, a 12B model with 128k context length context length The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. ollama run mistral-nemo. Adds NuExtract, a 3.8B Phi-3-based model fine-tuned for structured information extraction. ollama create now validates templates at creation time, catching errors before deployment.
3 details in Everything
Ollama v0.2.6 adds Mathstral, a 7B math reasoning model from Mistral AI.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.2.6 https://github.com/ollama/ollama.git
$ git checkout v0.2.6
Ollama v0.2.4 enables loading models larger than system memory when disk space is available.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.2.4 https://github.com/ollama/ollama.git
$ git checkout v0.2.4
Ollama v0.2.0 adds concurrent request handling, multi-model loading, and three new models including Gemma 2.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.2.0 https://github.com/ollama/ollama.git
$ git checkout v0.2.0
└──▷ TRY IT
Check which models are currently loaded into GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
New ollama ps subcommand shows which models are currently loaded, along with size, processor utilization, and expiry time. Supports serving multiple simultaneous requests in parallel using only a small amount of additional memory per request — enabling concurrent chat sessions, code completion, document processing, and multi-agent workloads. Supports loading multiple different models into GPU memory at the same time, with automatic load/unload based on request demand and available GPU memory — enabling RAG RAG Retrieval-augmented generation: looking up relevant documents and adding them to the prompt before the model answers. It grounds responses in private or current data without retraining. embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. Adds GLM-4, a multilingual general-purpose model competitive with Llama 3, to the Ollama model library. Adds CodeGeeX4, a code-completion and AI software development model, to the Ollama model library.
+2 more
show less
Adds Gemma 2, Google's next-generation model with improved output quality, to the Ollama model library. Returns a descriptive error on Linux when the system lacks sufficient memory to load a model.
7 details in Everything
Ollama v0.1.48 adds expanded model info display via /show info in ollama run
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.48 https://github.com/ollama/ollama.git
$ git checkout v0.1.48
└──▷ TRY IT
Inspect detailed model information while in an interactive session.
copy
$ ollama run gemma2
>>> /show info
The /show info command in ollama run now displays additional model information.
1 detail in Everything
Ollama v0.1.47 adds support for Google Gemma 2 models (9B and 27B).
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.47 https://github.com/ollama/ollama.git
$ git checkout v0.1.47
└──▷ TRY IT
Pull and run the Gemma 2 27B model locally for inference.
copy
$ ollama run gemma2:27b
Ollama v0.1.45 expands ollama show with full model metadata and adds DeepSeek-Coder-V2 support.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.45 https://github.com/ollama/ollama.git
$ git checkout v0.1.45
└──▷ TRY IT
Inspect a local model's context length context length The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
copy
$ ollama show llama3
Expands the api/show endpoint to return extensive model metadata. Reports GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. ollama serve output for improved observability. Enhanced GPU discovery and multi-GPU support with concurrency. Faster model loading on Windows with CUDA CUDA NVIDIA's platform and programming interface for running general computation on its GPUs. Nearly every AI training and inference stack is built on it first. GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
4 details in Everything
Ollama v0.1.42 adds Qwen 2 models, faster pulls, auto prompt-template detection, and Electron/Tauri app access.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.42 https://github.com/ollama/ollama.git
$ git checkout v0.1.42
ollama create now automatically detects prompt templates for popular model architectures including Llama, Gemma, and Phi.ollama pull is now faster when it detects a model is already partially or fully downloaded.Ollama can now be accessed from local apps built with Electron and Tauri, and from apps developed in local HTML files.
3 details in Everything
Ollama v0.1.40 adds Codestral, IBM Granite Code, and DeepSeek V2 models to the library.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.40 https://github.com/ollama/ollama.git
$ git checkout v0.1.40
Adds codestral to the Ollama model library — Mistral AI's first code model, designed for code generation tasks. Adds granite-code to the Ollama model library in 3b and 8b parameter sizes from IBM. Adds deepseek-v2 to the Ollama model library — a Mixture-of-Experts Mixture-of-Experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
3 details in Everything
Ollama v0.1.39 adds Llama 3 Safetensors Safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face. flash-attention flash-attention A faster, memory-efficient way to compute attention on a GPU by working in on-chip memory tiles. Most training and serving stacks use it to handle long contexts. OLLAMA_NOHISTORY support
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.39 https://github.com/ollama/ollama.git
$ git checkout v0.1.39
└──▷ TRY IT
Import and quantize quantize Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
copy
$ ollama create --quantize q4_0 -f Modelfile my-llama3
Adds OLLAMA_NOHISTORY=1 environment variable to disable shell history when using ollama run. Adds experimental OLLAMA_FLASH_ATTENTION=1 environment variable flag for ollama serve to improve token generation speed on Apple Silicon Macs and NVIDIA GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Adds --quantize flag to ollama create (e.g. --quantize q4_0) enabling import and quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. ollama create now supports creating models from I-Quant GGUF GGUF The single-file model format used by llama.cpp and Ollama, holding the weights, usually quantized, together with the tokenizer and settings needed to run the model locally. Adds Ctrl+W keyboard shortcut to ollama run.
+5 more
show less
Adds Cohere Aya 23 (aya), a multilingual LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Adds Mistral 7B v0.3 (mistral:v0.3) with initial function calling function calling A model's ability to answer with a structured request to run a named function or API, which the application executes and returns. It is how agents take actions. Adds Phi-3 Medium (phi3:medium), a 14B-parameter open model by Microsoft, to the model library. Adds Phi-3 Mini 128K (phi3:mini-128k) and Phi-3 Medium 128K (phi3:medium-128k) with 128K context window context window The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. Adds IBM Granite Code (granite-code), a family of open foundation models for code intelligence, to the model library.
10 details in Everything
Ollama v0.1.38 adds ollama ps to inspect loaded models and /clear to reset chat session history.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.38 https://github.com/ollama/ollama.git
$ git checkout v0.1.38
└──▷ TRY IT
Check which models are currently loaded into memory and how much GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. CPU CPU The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware.
Reset the conversation context mid-session without exiting, useful when starting a new topic in a long ollama run chat.
copy
$ ollama run llama3
>>> /clear
Adds ollama ps command to display currently loaded models, their memory footprint (in GB/MB), and processor usage (GPU vs CPU percentages). Adds /clear command inside ollama run sessions to reset chat history without ending the session. Adds Falcon 2, an 11B-parameter causal decoder-only model trained on 5T tokens, available via ollama pull falcon2. Adds Yi 1.5 (Apache 2.0) in 6B (yi:6b), 9B (yi:9b), and 34B (yi:34b) sizes.
4 details in Everything
Ollama v0.1.35 adds on-the-fly quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. --quantize flag and a new done_reason field in API responses.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.35 https://github.com/ollama/ollama.git
$ git checkout v0.1.35
└──▷ TRY IT
Quantize Quantize Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
copy
$ ollama create -f Modelfile --quantize q4_0 mymodel
Adds --quantize / -q flag to ollama create to quantize float16float16 Half-precision, a 16-bit floating-point format that uses less memory than 32-bit numbers. It has a narrower numeric range than BF16, so training often needs extra safeguards against overflow. float32 models (from GGUF GGUF The single-file model format used by llama.cpp and Ollama, holding the weights, usually quantized, together with the tokenizer and settings needed to run the model locally. q4_0. Adds done_reason field to API responses describing why generation stopped. Adds Llama 3 ChatQA model (llama3-chatqa), an NVIDIA model optimised for conversational QA and retrieval-augmented generation retrieval-augmented generation Retrieval-augmented generation: looking up relevant documents and adding them to the prompt before the model answers. It grounds responses in private or current data without retraining.
3 details in Everything
Ollama v0.1.34 adds five new models including multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.34 https://github.com/ollama/ollama.git
$ git checkout v0.1.34
└──▷ TRY IT
Run the new multimodal Llava Llama 3 model to analyze an image from the command line.
copy
$ ollama run llava-llama3 Pull and run the new StarCoder2 15B Instruct model for code generation tasks.
copy
$ ollama run starcoder2:15b-instruct
Adds Llava Llama 3 (llava-llama3), a high-performing multimodal LLaVA model fine-tuned from Llama 3 Instruct. Adds Llava Phi 3 (llava-phi3), a small multimodal LLaVA model fine-tuned from Phi 3. Adds StarCoder2 15B Instruct (starcoder2:15b-instruct), an instruction-tuned variant of the StarCoder2 code model. Adds CodeGemma 1.1 (codegemma), an updated release of Google's CodeGemma model. Adds StableLM2 12B (stablelm2:12b), a new 12B parameter version of Stability AI's StableLM 2 model.
+1 more
show less
Updates Moondream 2 (moondream) with improved runtime parameters for better response quality.
6 details in Everything
Ollama v0.1.33 adds Llama 3 and experimental concurrency via OLLAMA_NUM_PARALLEL and OLLAMA_MAX_LOADED_MODELS
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.33 https://github.com/ollama/ollama.git
$ git checkout v0.1.33
└──▷ TRY IT
Run the Ollama server with parallel request handling and multiple loaded models to serve concurrent users without queuing.
copy
$ OLLAMA_NUM_PARALLEL=4 OLLAMA_MAX_LOADED_MODELS=4 ollama serve
Adds OLLAMA_NUM_PARALLEL environment variable to handle multiple simultaneous requests for a single model (experimental). Adds OLLAMA_MAX_LOADED_MODELS environment variable to load multiple models into memory simultaneously (experimental). Adds Llama 3 (Meta), Phi 3 Mini (Microsoft 3.8B), Moondream (edge vision-language model vision-language model A vision-language model that works with images and text together, such as answering questions about a picture or describing what it contains.
3 details in Everything
Ollama v0.1.32 adds five new models including WizardLM 2 and Mixtral 8x22B, plus smarter GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. CPU CPU The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.32 https://github.com/ollama/ollama.git
$ git checkout v0.1.32
└──▷ TRY IT
Run the new WizardLM 2 7B model for fast, high-quality chat without pulling it separately first.
copy
$ ollama run wizardlm2:7b Pull the Snowflake Arctic Embed model to generate text embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. RAG RAG Retrieval-augmented generation: looking up relevant documents and adding them to the prompt before the model answers. It grounds responses in private or current data without retraining.
copy
$ ollama pull snowflake-arctic-embed
Adds wizardlm2:8x22b and wizardlm2:7b (WizardLM 2 from Microsoft AI) with improved performance on complex chat, multilingual, reasoning, and agent use cases. Adds snowflake-arctic-embed, a suite of text embedding models embedding models Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. Adds command-r-plus, a large language model large language model A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Adds dbrx, a 132B open general-purpose LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Adds mixtral:8x22b, Mistral AI's new Mixture of Experts Mixture of Experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
+2 more
show less
On macOS, Ollama now splits models that exceed available VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory. Improves VRAM utilization to reduce out-of-memory out-of-memory An out-of-memory condition: a program requests more memory than the system or device can provide. The allocation fails or the operating system may stop the process.
7 details in Everything
Ollama v0.1.31 adds embedding model embedding model Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. REST API REST API A web API design style where clients interact with resources over HTTP using standard methods like GET, POST, and DELETE, giving cyber tools a predictable, language-agnostic way to integrate with external services. RAG RAG Retrieval-augmented generation: looking up relevant documents and adding them to the prompt before the model answers. It grounds responses in private or current data without retraining.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.31 https://github.com/ollama/ollama.git
$ git checkout v0.1.31
Adds embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. ollama-python), and JavaScript library (ollama-js), enabling retrieval-augmented generation retrieval-augmented generation Retrieval-augmented generation: looking up relevant documents and adding them to the prompt before the model answers. It grounds responses in private or current data without retraining. Adds Qwen 1.5 32B (qwen:32b), a multilingual model competitive with larger models. Adds StarlingLM Beta (starling-lm:beta), a 7B model with Apache 2.0 license. Adds DolphinCoder StarCoder 7B (dolphincoder:7b), an uncensored coding-focused variant based on StarCoder2. Adds StableLM 1.6 Chat (stablelm2:chat), an instruction-tuned version of StableLM 1.6.
5 details in Everything
Ollama v0.1.30 adds Command R and mxbai-embed-large models plus improved AMD MI300/MI300X accelerator support.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.30 https://github.com/ollama/ollama.git
$ git checkout v0.1.30
└──▷ TRY IT
Pull and run Command R for long-context or conversational tasks.
copy
$ ollama run command-r Pull the mxbai-embed-large embedding model embedding model Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
copy
$ ollama run mxbai-embed-large
Adds command-r to the model library: a large language model large language model A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Adds mxbai-embed-large to the model library: a state-of-the-art large embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. Improved support for AMD MI300 and MI300X accelerators.
3 details in Everything
└──▷ BREAKING ON UPGRADE
! Model data filenames in ~/.ollama/models/blobs are renamed from using : to - as a separator. The change is applied automatically on upgrade, but downgrading to v0.1.29 or lower on Linux or macOS requires manually reverting the filenames: find ~/.ollama/models/blobs -type f -exec bash -c 'mv "$0" "${0//-/:}"' {} \;
Ollama v0.1.29 adds AMD GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Safetensors Safetensors A model weight file format that stores only tensors, unlike Python pickle files that can run code when loaded. It is the default on Hugging Face.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.29 https://github.com/ollama/ollama.git
$ git checkout v0.1.29
└──▷ TRY IT
Import a locally downloaded Safetensors model into Ollama for inference without converting it first.
copy
$ FROM /path/to/safetensors-model-dir
# Save as a Modelfile, then:
ollama create my-model -f Modelfile
Adds experimental Safetensors model import via the FROM <directory with safetensors model> directive in a Modelfile. Shows documentation for supported environment variables when running ollama <command> -h. Adds AMD GPU acceleration support (preview) on Windows, Linux, and Docker, covering AMD Radeon RX, Radeon PRO, and Instinct families including MI300X, 7900 XTX, W7900, and more.
3 details in Everything
Ollama v0.1.28 adds StarCoder2 and DolphinCoder models with improved LLaVA 1.6 support.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.28 https://github.com/ollama/ollama.git
$ git checkout v0.1.28
└──▷ TRY IT
Pull and run the new StarCoder2 code model in a size that fits your hardware.
copy
$ ollama run starcoder2:7b Pull and run DolphinCoder for conversational code-generation tasks.
copy
$ ollama run dolphincoder
Adds starcoder2 model (3B, 7B, and 15B parameter variants), a transparently trained open code LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Adds dolphincoder model, a code-focused chat model based on StarCoder2 15B. Improves support for llava 1.6 models. Vision models such as llava now respond better to text-only prompts.
4 details in Everything
Ollama v0.1.27 adds support for Google's Gemma family of models in 2b and 7b parameter sizes.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.27 https://github.com/ollama/ollama.git
$ git checkout v0.1.27
└──▷ TRY IT
Run Google's Gemma 2b model locally for lightweight inference tasks.
copy
$ ollama run gemma:2b
Ollama v0.1.26 adds support for bert and nomic-bert embedding models embedding models Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.26 https://github.com/ollama/ollama.git
$ git checkout v0.1.26
Ollama v0.1.25 brings native Windows preview with GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.25 https://github.com/ollama/ollama.git
$ git checkout v0.1.25
Setting keep_alive to a negative value (e.g. -1) now keeps a model loaded indefinitely, enabling persistent in-memory model serving. Ollama is now available natively on Windows (preview), with built-in GPU acceleration, access to the full model library, and the Ollama API including OpenAI compatibility.
2 details in Everything
Ollama v0.1.24 adds OpenAI Chat Completions API compatibility and the Qwen 1.5 model family.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.24 https://github.com/ollama/ollama.git
$ git checkout v0.1.24
└──▷ TRY IT
Use an existing OpenAI-compatible client or curl against your local Ollama instance instead of OpenAI's servers.
copy
$ curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "llama2",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Hello!"}
]
}'
Adds POST /v1/chat/completions endpoint providing initial OpenAI Chat Completions API compatibility, enabling existing OpenAI-compatible clients and SDKs to point at a local Ollama instance without code changes. Adds Qwen 1.5 model family (qwen) from Alibaba Cloud, spanning 0.5B to 72B parameter sizes, available via the Ollama model library.
2 details in Everything
Ollama v0.1.23 adds keep_alive API parameter, LLaVA 1.6 vision models, and broader Nvidia GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.23 https://github.com/ollama/ollama.git
$ git checkout v0.1.23
└──▷ TRY IT
Unload a model from memory immediately after a response to free VRAM VRAM Memory attached to a graphics processor. Models, intermediate calculations and caches must fit here or be moved to slower memory.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "mistral", "prompt": "Summarize this document.", "keep_alive": "0"}' Keep a model pinned in memory indefinitely so repeated requests incur no reload latency.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "mistral", "prompt": "Hello", "keep_alive": "-1"}'
Adds keep_alive parameter to the /api/generate (and /api/chat) API to control how long a model stays loaded in memory — accepts positive durations (e.g. 20m, 1hr, 30), 0 to unload immediately, or a negative value (e.g. -1) to keep loaded indefinitely; default is 5 minutes. Enables passing image paths to ollama run when running multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats. ollama serve now prints the prompt when the OLLAMA_DEBUG=1 environment variable is set.ollama run now allows sending messages without images to multimodal models.Updates the LLaVA model family to version 1.6, adding a new llava:34b variant alongside the existing llava (7B) and llava:13b, with support for up to 4x higher image resolution and improved text/chart/diagram reasoning.
+1 more
show less
Adds GPU support for Nvidia compute capability 5 devices, including GeForce GTX 900/700-series, Quadro M-series, Tesla M60/M40, and NVS 810.
6 details in Everything
Ollama v0.1.22 adds Stable LM 2, a 1.6B small language model.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.22 https://github.com/ollama/ollama.git
$ git checkout v0.1.22
└──▷ TRY IT
Pull and run the new Stable LM 2 1.6B model for lightweight, fast local inference.
copy
$ ollama run stablelm2
Ollama v0.1.21 adds conversation save/load, MESSAGE Modelfile command, Python/JS libraries, and broader CPU CPU The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.21 https://github.com/ollama/ollama.git
$ git checkout v0.1.21
└──▷ TRY IT
Seed a model with Chain-Of-Thought examples at build time so every session starts with pre-loaded conversation history.
copy
$ # Modelfile
FROM llama2
SYSTEM You are a helpful assistant.
MESSAGE user Is Toronto in Canada?
MESSAGE assistant yes
MESSAGE user Is Sacramento in Canada?
MESSAGE assistant no
# Then build and run:
ollama create -f Modelfile yesno
ollama run yesno
Adds /save <model> and /load <model> commands inside ollama run to persist and restore conversations and model settings (including /set parameter and /set system changes) as a named model. Adds MESSAGE Modelfile command to pre-seed conversation history when building a model with ollama create, enabling techniques like Chain-Of-Thought prompting. Publishes first-release official Python (ollama-python) and JavaScript (ollama-js) client libraries for Ollama. Extends CPU support to processors without AVX instructions, enabling Ollama to run in virtual machines, Rosetta, and GitHub Actions GitHub Actions GitHub's built-in CI/CD system that runs automated workflows triggered by repository events, letting cyber tools automate testing, building, and releasing without managing separate pipeline infrastructure. Delivers ~10% model inference speed boost on CPUs CPUs The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware.
+3 more
show less
Adds GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Adds four new models to the library: Qwen (1.8B–72B), DuckDB-NSQL (text-to-SQL for DuckDB), Stable Code, and Nous Hermes 2 Mixtral. Improves Nvidia GPU detection, especially under WSL WSL Windows Subsystem for Linux, a Microsoft-built compatibility layer that lets a Linux environment run directly on Windows without a virtual machine, giving cyber tools access to Linux binaries, filesystems, and syscalls on a Windows host.
8 details in Everything
Ollama v0.1.20 adds MegaDolphin 120B and updates OpenChat and Dolphin Mistral models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.20 https://github.com/ollama/ollama.git
$ git checkout v0.1.20
└──▷ TRY IT
Pull and run the new MegaDolphin 120B model locally.
copy
$ ollama run megadolphin
Adds megadolphin to the model library: a 120B parameter version of the Dolphin model. Updates openchat to version 3.5-0106. Updates dolphin-mistral to the latest DPO DPO Direct Preference Optimization, a simpler alternative to RLHF that tunes a model straight from pairs of preferred and rejected answers, with no separate reward model. Adds correct multi-GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
4 details in Everything
Ollama v0.1.19 adds LLaMa-Pro model support, expandable context windows context windows The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. num_ctx, and broader OLLAMA_ORIGINS URL support.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.19 https://github.com/ollama/ollama.git
$ git checkout v0.1.19
└──▷ TRY IT
Expand Mistral's context window context window The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "mistral", "prompt": "Summarize this document:", "options": {"num_ctx": 32678}}' Set a 32K context window interactively in an ollama run session before sending a long prompt.
copy
$ ollama run mistral
/set parameter num_ctx 32678
Sets context window size via num_ctx in /set parameter num_ctx (CLI) or the options.num_ctx field in the POST /api/generate JSON body — enabling up to 32K context with models like Mistral. OLLAMA_ORIGINS environment variable now accepts browser extension URLs, widening permitted request origins.Adds the LLaMa-Pro model to the library — an 8B expansion of LLaMa by Tencent specializing in language, programming, and mathematics. Offloads more processing to the GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. Enables larger models such as mixtral to run on Macs with less memory.
5 details in Everything
Ollama v0.1.18 adds four new models and a keyboard shortcut help command in the CLI.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.18 https://github.com/ollama/ollama.git
$ git checkout v0.1.18
└──▷ TRY IT
Run the new WizardCoder 33B model for state-of-the-art code generation tasks.
copy
$ ollama run wizardcoder:33b
Adds /? shortcuts help command inside ollama run to list available keyboard shortcuts. Adds TinyLlama (1.1B) to the model library: ollama run tinyllama. Adds OpenHermes 2 (7B, Mistral fine-tune) to the model library: ollama run openhermes. Adds WizardCoder 33B code-generation model to the model library: ollama run wizardcoder:33b. Adds Dolphin Phi (2.7B, uncensored, Microsoft Phi-based) to the model library: ollama run dolphin-phi.
+1 more
show less
Ollama now automatically pulls new models when an older ggml format model is run, removing the need to manually fetch updated versions.
6 details in Everything
└──▷ BREAKING ON UPGRADE
! Custom ggml format models referenced in a Modelfile will no longer work as before — they must be imported as GGUF GGUF The single-file model format used by llama.cpp and Ollama, holding the weights, usually quantized, together with the tokenizer and settings needed to run the model locally.
Ollama v0.1.17 adds Microsoft Phi-2, Solar, and OpenChat models plus a guaranteed message object in /api/chat responses.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.17 https://github.com/ollama/ollama.git
$ git checkout v0.1.17
└──▷ TRY IT
Run a multi-turn conversation with the new Phi-2 model locally.
Query Phi-2 via the chat API for integration into an application pipeline.
copy
$ curl http://localhost:11434/api/chat -d '{"model": "phi", "messages": [{"role": "user", "content": "why is the sky blue?"}]}'
Adds phi model support via ollama run phi — Microsoft's Phi-2 2.7B model with reasoning and language understanding comparable to 13B models, including a default multi-turn conversation prompt template. /api/chat now always returns a message object in every response, even when content is an empty string.Adds solar model (ollama run solar) — a 10.7B model designed for single-turn conversation. Updates openchat model to OpenChat-3.5-1210, a 7B model with improved coding task performance. Updates wizard-math model to WizardMath v1.1, a 7B math reasoning model now based on Mistral.
5 details in Everything
Ollama v0.1.16 adds Mixtral and Mixture of Experts Mixture of Experts A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token. MoE MoE A model design that splits its layers into many expert sub-networks and routes each token to only a few. The model holds many parameters but spends the compute of a much smaller one per token.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.16 https://github.com/ollama/ollama.git
$ git checkout v0.1.16
└──▷ TRY IT
Run the Mixtral MoE model locally for high-quality inference on a machine with at least 48 GB of memory.
copy
$ ollama run mixtral
Adds support for the Mixtral model via ollama run mixtral (requires at least 48 GB of memory), enabling use of a high-quality open-weights mixture of experts model. Adds support for Mixture of Experts (MoE) architecture broadly, enabling other MoE-based models beyond Mixtral. Adds Dolphin Mixtral (jmorgan/dolphin-mixtral), an uncensored fine-tuned MoE model optimized for coding tasks.
3 details in Everything
Ollama v0.1.15 adds multimodal multimodal Able to work with more than one type of input or output, such as text, images and audio. A multimodal model can combine information across those formats.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.15 https://github.com/ollama/ollama.git
$ git checkout v0.1.15
└──▷ TRY IT
Query a vision model about an image from the command line without writing any code.
copy
$ ollama run llava
>>> What does the text in this image say? /Users/you/Downloads/image.png Send an image to a multimodal model via the Generate API for automated pipelines or scripting.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "llava", "prompt": "What is in this picture?", "images": ["<base64-encoded-png-or-jpeg>"]}' Inspect a model's format, architecture families, parameter size, and quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware.
copy
$ curl http://localhost:11434/api/show -d '{"name": "llava"}'
Adds images parameter to the Generate API (POST /api/generate) accepting a list of base64-encoded png or jpeg images up to 100MB each, enabling vision-capable models like LLaVA. Adds images field to user role messages in the Chat API (POST /api/chat), allowing image inputs in multi-turn conversations. Extends POST /api/show to return a details object with format, families, parameter_size, and quantization_level fields for the requested model. Adds multimodal image input to ollama run: include a full path to a png or jpeg file directly in the prompt to pass images to vision models. Adds two new multimodal models to the library: llava (7B and 13B) and bakllava (Mistral 7B + LLaVA architecture).
5 details in Everything
Ollama v0.1.14 adds a Chat API with message history, a version endpoint, and two new models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.14 https://github.com/ollama/ollama.git
$ git checkout v0.1.14
└──▷ TRY IT
Send a multi-turn conversation to a local model, preserving system prompt and prior user turns in a single request.
copy
$ curl http://localhost:11434/api/chat -d '{"model": "mistral", "messages": [{"role": "system", "content": "You are a helpful assistant that answers concisely."}, {"role": "user", "content": "why is the sky blue?"}]}'
New POST /api/chat endpoint accepts a messages array with role and content fields, enabling multi-turn chat with full conversation history passed per request. New GET /api/version endpoint for programmatically checking the running Ollama version. Adds version-mismatch warning when connecting to a remote instance via OLLAMA_HOST. Adds StableLM Zephyr (stablelm-zephyr), a lightweight chat model designed for low-end hardware. Adds Magicoder (magicoder), a family of 7B models trained on 75K synthetic OSS OSS Open-source software: code whose licence permits anyone to inspect, modify, and redistribute it. Cyber tools build on it to reuse auditable libraries without licensing fees.
5 details in Everything
Ollama v0.1.13 adds in-session system prompt and parameter tuning via /set, plus three new models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.13 https://github.com/ollama/ollama.git
$ git checkout v0.1.13
└──▷ TRY IT
Override the system prompt mid-session to repurpose a running model without reloading it.
copy
$ /set system You are a concise assistant that replies only in bullet points. Increase context window context window The maximum amount of text, counted in tokens, a model can take in and produce in one request. Everything the model knows about the task has to fit inside it. ollama run session.
copy
$ /set parameter num_ctx 16384
/set parameter temperature 0.2
/set parameter seed 1048
Adds /set system <system prompt> command inside ollama run to set the system prompt interactively during a session. Adds /set parameter <parameter> <value> command inside ollama run to tune inference parameters (e.g. num_ctx, temperature, seed) without restarting. Adds three new models to the Ollama library: starling-lm (RLHF RLHF Reinforcement learning from human feedback: tuning a model toward the responses people rate higher, using a reward model trained on those ratings. meditron (Llama 2 adapted for medical domain), and deepseek-llm (2-trillion-token bilingual LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Improves ollama pull progress bar with a simpler design showing more consistent download speed and remaining time.
4 details in Everything
Ollama v0.1.12 adds Yi Chat 34B and improves multi-line prompt handling in the CLI.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.12 https://github.com/ollama/ollama.git
$ git checkout v0.1.12
Adds Yi Chat (the chat variant of the Yi 34B model) to the Ollama model library, available via ollama run yi. Supports multi-line prompts delimited by """ and improved paste functionality in ollama run. Adds Option (Alt) + Backspace word-deletion keybinding in the ollama run interactive prompt.
3 details in Everything
Ollama v0.1.11 adds Orca 2, DeepSeek Coder, and Alfred models plus GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.11 https://github.com/ollama/ollama.git
$ git checkout v0.1.11
Adds q5_0 and q5_1 quantized quantized Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. Adds Orca 2 model (orca2), a Llama 2 fine-tune optimized for reasoning tasks. Adds DeepSeek Coder model (deepseek-coder), a code-focused model available in 1.3B, 6.7B, and 33B parameter sizes. Adds Alfred model (alfred), a conversational model supporting both chat and instruct use cases.
4 details in Everything
Ollama v0.1.10 adds JSON mode JSON mode Forcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. ollama run, stdin prompt piping, and remote model builds via OLLAMA_HOST.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.10 https://github.com/ollama/ollama.git
$ git checkout v0.1.10
└──▷ TRY IT
Feed a local file directly into a model for analysis without copying its contents manually.
copy
$ head -30 README.md | ollama run codellama 'how do I install Ollama on Linux?'
Adds --format json flag to ollama run to enable JSON output mode from the command line. Adds /set format json in-session command to switch an active ollama run chat session to JSON mode. ollama create now respects OLLAMA_HOST to build models against a remote Ollama instance.Enables piping prompts into ollama run via standard input, allowing shell pipelines like head -30 README.md | ollama run codellama '<question>'. Adds three new models to the library: OpenChat (ollama run openchat), Neural-chat (ollama run neural-chat), and Goliath (ollama run goliath).
5 details in Everything
Ollama v0.1.9 adds JSON mode JSON mode Forcing a model's response to match a given format or JSON schema, so software can parse it reliably instead of hoping the text is well formed. /api/generate, plus a new bilingual Yi model.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.9 https://github.com/ollama/ollama.git
$ git checkout v0.1.9
└──▷ TRY IT
Force structured JSON output from a model — useful when piping responses into a parser or downstream tool.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "llama2", "prompt": "List the top 3 open ports on a typical web server.", "format": "json"}' Send a fully pre-formatted prompt without Ollama applying any chat template — useful when you control the prompt structure yourself.
copy
$ curl http://localhost:11434/api/generate -d '{"model": "llama2", "prompt": "[INST] Summarize this CVE. [/INST]", "raw": true}'
Adds format parameter to POST /api/generate — set it to json to force models to always return valid JSON (JSON mode). Adds raw parameter to POST /api/generate — set {"raw": true} to bypass Ollama's prompt templating entirely (raw mode). Adds the Yi bilingual model (English and Chinese) to the Ollama library, available via ollama pull yi.
3 details in Everything
Ollama v0.1.8 adds five new models and dramatically faster push speeds up to 1 GB/s for large models.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.8 https://github.com/ollama/ollama.git
$ git checkout v0.1.8
Adds codebooga to the model library: a high-performing code instruct model created by merging two existing code models. Adds dolphin2.2-mistral to the model library: a Mistral-based instruct-tuned model fine-tuned for improved conversation and empathy. Adds mistrallite to the model library: a Mistral fine-tune with enhanced long-context processing capabilities. Ollama now honours large context sizes on models such as codellama and mistrallite. ollama push is now dramatically faster: 7B models push at up to ~100 MB/s and 70B+ models at up to 1 GB/s when network permits.
5 details in Everything
Ollama v0.1.6 adds two new models, multi-line paste in ollama run, and configurable model storage via OLLAMA_MODELS.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.6 https://github.com/ollama/ollama.git
$ git checkout v0.1.6
└──▷ TRY IT
Set a custom system prompt for a chat model directly in a Modelfile now that all instruct/chat models support the SYSTEM command.
copy
$ FROM llama2
SYSTEM You are a concise assistant that responds only in bullet points.
Adds OLLAMA_MODELS environment variable to configure where Ollama stores models. OLLAMA_HOST now defaults to port 443 when https:// is specified and port 80 when http:// is specified, removing the need to manually append ports.Enables pasting multi-line strings in ollama run. All chat and instruct library models now support setting the system parameter via the SYSTEM command in the Modelfile . Adds Dolphin 2.1 Mistral (dolphin2.1-mistral), an instruct-tuned model based on Mistral trained on a dataset filtered to remove alignment and bias.
+1 more
show less
Adds Zephyr Beta (zephyr), a Mistral-based model with chat performance reported to meet or exceed Llama 2 70b in several categories.
6 details in Everything
Ollama v0.1.4 adds OpenHermes 2 Mistral, faster model switching, and unicode support for starcoder/sqlcoder/falcon.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.4 https://github.com/ollama/ollama.git
$ git checkout v0.1.4
└──▷ TRY IT
Pull the updated starcoder build to gain unicode support after the v0.1.4 upgrade.
copy
$ ollama pull starcoder
Adds ollama pull starcoder (and sqlcoder, falcon) to get updated builds with full unicode support — existing pulls must be re-pulled. Faster model switching: models stay loaded between requests when switching parameters (e.g. temperature) or system prompts, eliminating reload overhead. Adds OpenHermes 2 Mistral to the model library (ollama pull openhermes2-mistral), a Mistral fine-tune trained on 900,000+ instructions with strong multi-turn chat performance. ollama serve now prints the current Ollama version on startup.ollama run now surfaces descriptive runtime errors (e.g. insufficient memory) instead of silent failures.
+1 more
show less
New documentation guide for importing existing models (GGUF GGUF The single-file model format used by llama.cpp and Ollama, holding the weights, usually quantized, together with the tokenizer and settings needed to run the model locally.
6 details in Everything
Ollama v0.1.2 adds non-streaming API responses, HTTP/HTTPS HTTPS HTTP carried over an encrypted TLS connection. It protects traffic in transit and lets a browser verify the server's certificate.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.2 https://github.com/ollama/ollama.git
$ git checkout v0.1.2
└──▷ TRY IT
Retrieve a complete model response in one shot — useful in scripts or pipelines that expect a single JSON object rather than a streamed token sequence.
copy
$ curl -X POST http://localhost:11434/api/generate -d '{"model": "llama2", "prompt": "Why is the sky blue?", "stream": false}'
Adds stream: false parameter to API endpoints (e.g. /api/generate) to return a complete, non-streaming response in a single payload instead of a token stream. Supports HTTP and HTTPS proxies via the HTTP_PROXY=http://<proxy> and HTTPS_PROXY=https://<proxy> environment variables. Adds two new models to the library: zephyr (fine-tuned Mistral 7B) and mistral-openorca (Mistral 7B fine-tuned on the OpenOrca dataset). ollama pull download speeds improved from ~60 MB/s to over 1.5 GB/s (25x) on fast network connections, enabling practical use of large models over high-bandwidth links.q8_0, q5_0, q5_1, and f32 quantization quantization Storing a model's weights at lower numeric precision, such as 8 or 4 bits instead of 16, so it needs less memory and runs faster. It trades a small amount of accuracy for fitting on smaller hardware. GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
+1 more
show less
ollama create now displays feedback while reading model metadata.
6 details in Everything
Ollama v0.1.1 adds cancellable responses and clean session exit controls for ollama run.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.1 https://github.com/ollama/ollama.git
$ git checkout v0.1.1
Adds Ctrl+C to cancel in-progress responses during an ollama run session. Adds Ctrl+D and /bye as exit commands for ollama run sessions.
2 details in Everything
Ollama v0.1.0 adds Linux support with out-of-the-box Nvidia GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.1.0 https://github.com/ollama/ollama.git
$ git checkout v0.1.0
└──▷ TRY IT
Install Ollama on a Linux server or WSL WSL Windows Subsystem for Linux, a Microsoft-built compatibility layer that lets a Linux environment run directly on Windows without a virtual machine, giving cyber tools access to Linux binaries, filesystems, and syscalls on a Windows host.
copy
$ curl https://ollama.ai/install.sh | sh
Adds Linux support with Nvidia GPU acceleration enabled out-of-the-box, installable via curl https://ollama.ai/install.sh | sh. Supports multi-GPU cloud servers, automatically offloading as many model layers as the GPU(s) can handle for maximum performance. Supports WSL 2 with GPU acceleration. Runs on CPU CPU The main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware. GPUs GPUs A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
4 details in Everything
Ollama v0.0.20 adds IPv6 IPv6 The current version of the Internet Protocol, standardized by the IETF, using 128-bit addresses to replace the exhausted IPv4 space. Cyber tools target it because modern networks run dual-stack or IPv6-only.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.20 https://github.com/ollama/ollama.git
$ git checkout v0.0.20
└──▷ TRY IT
Pre-warm a model on a remote Ollama instance so the first real prompt returns immediately.
copy
$ curl http://<host>:11434/api/generate -d '{"model": "llama2", "prompt": ""}'
Adds api/client.py, a Python client library for accessing Ollama programmatically. OLLAMA_HOST now accepts IPv6 hostnames.Sending an empty prompt field to /api/generate pre-loads the model so the next request is faster. ollama run now loads models immediately on startup, reducing latency even on the first prompt.ollama run automatically pulls models that don't exist when connected to a remote Ollama instance.
+2 more
show less
ollama run gains hint text display and terminal-width word-wrapping for improved readability.Progress reporting improved for ollama pull and ollama push.
7 details in Everything
Ollama v0.0.19 adds GGUF GGUF The single-file model format used by llama.cpp and Ollama, holding the weights, usually quantized, together with the tokenizer and settings needed to run the model locally.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.19 https://github.com/ollama/ollama.git
$ git checkout v0.0.19
Supports importing and running GGUF file type models locally. Automatically cleans up unused local model data. Updated Docker image available via docker pull ollama/ollama.
3 details in Everything
Ollama v0.0.18 adds ollama show command to inspect model system prompts, parameters, templates, and Modelfiles.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.18 https://github.com/ollama/ollama.git
$ git checkout v0.0.18
└──▷ TRY IT
Inspect a model's system prompt, parameters, template, or full Modelfile before deploying it in a pipeline.
copy
$ ollama show --system orca-mini
ollama show --parameters codellama
ollama show --template llama2
ollama show --modelfile llama2
New ollama show command with --system, --parameters, --template, and --modelfile flags to inspect a model's system prompt, parameters, default prompt template, and Modelfile respectively. Adds a new sentiments example contributed by @technovangelist. Building from source now requires running go generate ./... to generate dependencies, with cmake as a new build prerequisite.
3 details in Everything
└──▷ BREAKING ON UPGRADE
! Building from source now requires running go generate ./... before go build ., and cmake must be installed — existing build workflows that skip this step will fail.
Ollama v0.0.17 adds multi-model removal and unique content-based IDs in ollama list.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.17 https://github.com/ollama/ollama.git
$ git checkout v0.0.17
└──▷ TRY IT
Clean up several models at once instead of running separate rm commands for each.
copy
$ ollama rm mario:latest orca-mini:3b
Supports removing multiple models in a single command: ollama rm mario:latest orca-mini:3b ollama list now displays a unique content-based ID for each model.
2 details in Everything
Ollama v0.0.16 adds version flags, 34B model support, and HTTPS HTTPS HTTP carried over an encrypted TLS connection. It protects traffic in transit and lets a browser verify the server's certificate. ollama run
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.16 https://github.com/ollama/ollama.git
$ git checkout v0.0.16
└──▷ TRY IT
Check the installed Ollama version from the command line.
Run a 34B model by referencing it with a full HTTPS path.
copy
$ ollama run https://example.com/models/codellama:34b
Adds ollama -v and ollama --version flags to check the installed Ollama version. Supports running 34B parameter models such as codellama via ollama run. Accepts model names or paths prefixed with https:// when using ollama run.
3 details in Everything
Ollama v0.0.15 adds remote host targeting, fine-tuned model adapters, smarter embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.15 https://github.com/ollama/ollama.git
$ git checkout v0.0.15
└──▷ TRY IT
Point your local Ollama CLI at a remote Ollama server to run inference without a local GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference.
copy
$ OLLAMA_HOST=192.168.1.50 ollama run llama2
Supports targeting remote Ollama hosts via OLLAMA_HOST=<host> ollama run llama2, enabling centralized model serving. Adds ADAPTER Modelfile instruction to ollama create for experimental creation of fine-tuned models. Adds context field to /api/generate API documentation, surfacing conversation context as a named API surface. Skips regenerating embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. EMBED Modelfile blocks when input files have not changed, reducing redundant computation. Reuses an already-loaded model for /api/embeddings requests when one is available, avoiding redundant model loads.
+4 more
show less
Parses memory requirements from ggml format models to validate GPU support before running. Launches public model library at ollama.ai/library, making curated models browsable and downloadable. Logs embedding eval timing for performance observability. Adds dockerit example tool for building and running applications in Docker containers.
9 details in Everything
Ollama v0.0.14 adds LangChain integration, embeddings embeddings Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. OLLAMA_HOST
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.14 https://github.com/ollama/ollama.git
$ git checkout v0.0.14
└──▷ TRY IT
Bind Ollama to a specific host and port so it is reachable across a local network, not just localhost.
copy
$ OLLAMA_HOST=0.0.0.0:11434 ollama serve Run the 70B Llama 2 model with Metal acceleration on a Mac with 64 GB+ RAM RAM Working memory that holds programs and data while a computer is running. Its contents usually disappear when power is removed.
copy
$ ollama run llama2:70b
New /api/embeddings endpoint enables embedding embedding Lists of numbers a model produces to represent the meaning of text, images or code, so similar items sit close together. They power semantic search and RAG. New EMBED instruction in the Modelfile for experimental embedding configuration. Adds LangChain integration for using Ollama as an LLM LLM A large language model, a neural network trained on large amounts of text to recognize patterns and generate language. Applications use it for tasks such as answering, summarizing and writing code. Adds llama2:70b model with Apple Metal GPU GPU A graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference. ollama run llama2:70b. Adds llama2-uncensored:70b model with Metal support, runnable via ollama run llama2-uncensored:70b.
+2 more
show less
Adds configurable rope frequency parameters for fine-grained model context control. ollama pull now supports concurrent pulls of the same model across multiple terminal windows.
7 details in Everything
Ollama v0.0.13 adds Intel Mac support, /show model inspector, --verbose load timing, and auto-start from CLI.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.13 https://github.com/ollama/ollama.git
$ git checkout v0.0.13
└──▷ TRY IT
Inspect the active model's metadata mid-session without leaving the prompt.
copy
$ ollama run llama2
>>> /show Send a multi-line prompt to a model in a single interactive input block.
copy
$ ollama run llama2
>>> """
Analyze the following log excerpt
and identify anomalies:
<paste log here>
"""
Adds --verbose flag to ollama run to display model load duration times. Adds /show command inside ollama run sessions to inspect the currently loaded model. Supports multi-line string input in ollama run using triple-quote (""") delimiters. Auto-starts the Ollama server when invoking the CLI without Ollama already running. Extends Intel Mac support, enabling Ollama to run on that platform for the first time.
+2 more
show less
Models now stay loaded in memory between messages automatically, accelerating series of prompts. Removes the update dialog in favor of a subtle tray-menu hint when a new version is ready.
7 details in Everything
Ollama v0.0.12 adds ollama cp for renaming models, k-quant support, and localhost CORS CORS A browser mechanism that lets a server declare which foreign origins may read its HTTP responses, enforced by the browser to block unauthorized cross-site requests from JavaScript.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.12 https://github.com/ollama/ollama.git
$ git checkout v0.0.12
Adds support for running k-quant models. ollama create now automatically pulls models referenced in a FROM instruction that are not available locally.Enables Apple Accelerate framework for performance improvements. Ollama's API now accepts requests from websites hosted on localhost (CORS support).
4 details in Everything
Ollama v0.0.11 adds ollama list and ollama rm subcommands plus checksum verification on ollama pull.
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.11 https://github.com/ollama/ollama.git
$ git checkout v0.0.11
└──▷ TRY IT
Audit local model storage to see what is taking up disk space before deciding what to remove.
Free up disk space by removing a model you no longer need.
copy
$ ollama rm llama2:13b
New ollama list subcommand displays all locally downloaded models with their size and last-modified time. New ollama rm subcommand deletes a locally stored model by name. The ollama pull subcommand now verifies downloaded model integrity against its checksum.
3 details in Everything
Ollama v0.0.3 adds larger prompt support, better error reporting, and moves logs to ~/.ollama/logs
└──▷ GET THIS VERSION
copy
$ git clone --branch v0.0.3 https://github.com/ollama/ollama.git
$ git checkout v0.0.3
Logs are now written to ~/.ollama/logs for easier debugging and audit. ollama run now supports a larger prompt, removing a previous size constraint.ollama run now prints better errors when a problem is encountered.Much faster performance across the board.
4 details in Everything
Choose a badge style and copy the code →