vMLX

No entries match these filters. Clear filters

1. v1.6.68 Sep 27, 2026 · release notes

   vMLX 1.6.68 adds native JANGH loading and panel exposes DSV4 thinking controls via new API/hook points.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.68 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.68

   • feat(runtime) introduces JANGH with legacy v2 compatibility

   1 detail in Everything

2. v1.6.65 Sep 24, 2026 · release notes

   vMLX 1.6.65 adds MiMo V2.6 runtime support for multimodalmultimodalAble 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. inputs and Markdown session export with model metadata.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.65 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.65

   • MiMo V2.6 runtime support for text, image, video and audio input, native reasoning and tool conversations
   • Markdown session export with model-provided reasoning, model identity, system prompts and recorded generation settings

   2 details in Everything

3. v1.6.56 Sep 8, 2026 · release notes

   vMLX 1.6.56 adds per-request image/video controls and typed strict-mode errors to the API dialects.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.56 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.56

   • Adds per-request image and video controls to the API, allowing fine-grained media handling for individual requests.
   • Introduces typed strict-mode error responses across all supported API dialects for better client-side validation.

   2 details in Everything

4. v1.6.35 Aug 22, 2026 · release notes

   vMLX v1.6.35 moves prompt caching to SSD by default with a disk-percentage budget and full-precision storage

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.35 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.35

   • Adds SSD-first caching as the default, turning the in-memory paged cache OFF by default to reduce memory use
   • Adds a session budget calculated as a percent of your disk (10% by default) instead of a fixed GB limit
   • Makes stored prefix caches full precision for every model family, eliminating TurboQuant approximation on read-back

   3 details in Everything

   └──▷ BREAKING ON UPGRADE

   • !The in-memory paged cache is now OFF by default; users must explicitly enable the RAMRAMWorking memory that holds programs and data while a computer is running. Its contents usually disappear when power is removed. tier via Session Settings to regain previous memory-based caching behavior
5. v1.6.29 Aug 15, 2026 · issue 047

   vMLX 1.6.29 adds hybrid prefix cache for 20x faster follow-ups and proves stable 100k-context multiturn with 99.9% cache reuse.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.29 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.29

   • Hybrid prefix cache delivers up to 20x faster follow-up turns on long documents (43.7k-token document follow-ups drop from 107s to 5s, byte-identical at temperature 0).
   • Chunked SSM re-derive ships enabled by default, restoring long-context cache reuse from 0% to 99.9% above 12.5k tokens on hybrid SSM model families.
   • 100k-context multiturn proven stable: 97.6k-token conversations retrieve planted facts with 99.9% cache reuse and no memory faults.
   • Per-family cache, parser, and generation-default hardening verified across 12 model families, including DSV4 Flash native composite caching, Gemma 4 mixed-SWA with audio, Qwen 3.6/3.8 hybrid lines with native MTP, and TurboQuant KV bundles.

   4 details in Everything

6. v1.6.28 Aug 14, 2026 · issue 045

   vMLX v1.6.28 adds native Qwen3.6-27B bundle support with automatic multi-token prediction, delivering a measured +33% throughput gain.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.28 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.28

   • Supports the Qwen3.6-27B bundle line (and the upcoming Qwen 3.8 line): the qwen3_coder tool-parser name now resolves to the XML-function parser, capability-only stamps route through the stock loader instead of the JANG codec, and the bundled multi-token prediction head is automatically constructed, quantizedquantizedStoring 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. per the stamp's per-module overrides, and engaged at the stamp's trained speculative depth — no flags required.
   • Delivers a measured 23.3 → 31 tokens/sec (+33%) on the 4-bit Qwen3.6-27B bundle with greedy output byte-identical whether the MTP head is on or off, and image/video probes answering correctly through both API and app.

   2 details in Everything

7. v1.6.27 Aug 11, 2026 · issue 042

   vMLX 1.6.27 adds Muse Glimmer 30B with vision/video, expands prefix cache to 262,144 tokens, and wires up Anthropic thinking.budget_tokens and Ollama think string levels.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.27 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.27

   └──▷ TRY IT
   Cap Muse Glimmer 30B reasoning budget via the Anthropic API so long chains do not run unbounded.
   copy

   $ curl http://localhost:8000/v1/messages \
     -H 'Content-Type: application/json' \
     -d '{
       "model": "local",
       "max_tokens": 4096,
       "thinking": {"type": "enabled", "budget_tokens": 2000},
       "messages": [{"role": "user", "content": "Explain zero-knowledge proofs."}]
     }'

   Drive reasoning effort from an Ollama client using the new think string level instead of a boolean.
   copy

   $ curl http://localhost:8000/api/chat \
     -H 'Content-Type: application/json' \
     -d '{
       "model": "local",
       "think": "high",
       "messages": [{"role": "user", "content": "Analyze this network capture for lateral movement."}]
     }'

   • Ollama think string levels ('low', 'medium', 'high') now engage thinking and select effort depth — previously they were silently ignored.
   • Prefix cache index expanded from ~64,000 to 262,144 tokens, with automatic migration of existing sessions, so long conversations no longer report zero reuse.
   • Adds full support for Muse Glimmer 30B: text tower, windowed vision tower, recipient-routed reasoning rail, and ATEM tool dialect — all three bundles (JANG_2D / 4M / 6M) supported.
   • Muse Glimmer 30B single- and multi-image prompts now correctly identify distinct objects in the same turn.
   • Muse Glimmer 30B video clips are now read in temporal order.

   +4 more show less

   • Chat Settings for Muse Glimmer 30B now exposes all four reasoning depths — Low / Medium / High / Extra High — replacing the previous Auto/On-only toggle.
   • Each reasoning depth in Muse Glimmer 30B maintains its own cache lineage, so switching back to a prior depth resumes that chain rather than starting over.
   • Paged RAMRAMWorking memory that holds programs and data while a computer is running. Its contents usually disappear when power is removed. is now consistently enabled for vision models between the app and command line, eliminating a silent cache-reuse penalty in the app.
   • Control markup (recipient markers) is now scrubbed from visible Muse answers on malformed or truncated streams, while reasoning text and tool-call bodies are left intact.

   9 details in Everything

8. v1.6.25 Aug 8, 2026 · issue 040

   vMLX 1.6.25 brings a 29–58% DeepSeek V4 prefillprefillThe 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. speedup, sub-3s retry TTFTTTFTTime to first token, the wait between submitting a request and receiving the first generated token. It includes delays such as queuing and processing the prompt. via KV rewind, and honest 413 errors instead of Metal OOMOOMAn 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. crashes on extreme context lengths.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.25 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.25

   • New indexed-attention prefill kernel (on by default) lifts DeepSeek V4 Flash prompt processingprompt processingThe 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. from 347 to ~449 tokens/s at 15k context (+29%) and from 199 to ~316 tokens/s at 40k context (+58%), while sustaining 33+ tokens/s decode at both lengths.
   • KV cacheKV cacheMemory 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. rewind on abort/retry: engine rewinds the live KV cache to the shared prefix instead of re-prefilling, dropping retry time-to-first-tokentime-to-first-tokenTime to first token, the wait between submitting a request and receiving the first generated token. It includes delays such as queuing and processing the prompt. at 15k context from 46s to 2.3s.
   • Two-pass answer path now reuses the first pass's KV cache, eliminating a full prompt re-prefill on the second pass.
   • KV pool quantizationquantizationStoring 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. keeps cache RAMRAMWorking memory that holds programs and data while a computer is running. Its contents usually disappear when power is removed. under 8 GB even near the model's maximum context.
   • Engine now advertises the true hardware memory ceiling at startup, shrinks prefill chunks adaptively under Metal memory pressure, and returns a clean 413 on over-ceiling requests instead of crashing with a Metal OOM; the flat request timeout that killed healthy very-long prefills is removed.

   5 details in Everything

9. v1.6.22 Aug 3, 2026 · issue 034

   vMLX 1.6.22 scales the MLX cache ceiling with installed memory, delivering ~11% faster decoding for DeepSeek V4 Flash.

   └──▷ GET THIS VERSION

   copy

   $ git clone --branch v1.6.22 https://github.com/jjang-ai/vmlx.git
   # already have the repo? check out this version:
   $ git checkout v1.6.22

10. v1.6.20 Aug 2, 2026 · issue 034

    vMLX v1.6.20 adds native DeepSeek V4 Flash runtime support with DSML tooling, pool-quantizationquantizationStoring 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., and enhanced cache reporting.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.6.20 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.6.20

    • Adds native DeepSeek V4 Flash (DSV4) session support with native reasoning, DSML tool support, sampling guidance, cache-tier capabilities, and pool-quantization behavior.
    • Server and Chat settings now distinguish native DSV4 compiled decode and pooled-cache state from generic TurboQuant and unsupported whole-model cache modes, with explicit Activation-QAT control and effective-state reporting.
    • Native DSV4 cache handling preserves short append checkpoints, lossless L2 writes, valid eviction ancestry, partial-tail replay, SSD-only operation, and RAMRAMWorking memory that holds programs and data while a computer is running. Its contents usually disappear when power is removed.-to-SSD refault.
    • Health output now distinguishes the served model from the currently loaded model.
    • Explicit visible-final instructions after tool usetool useA 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. now enter a no-more-tools continuation pass, with terminal abort cleanup drained before the runtime returns to idle.

    +2 more show less

    • Electron settings warn when a DSV4 top-p value differs from bundle guidance and preserve remote model tool contracts.
    • Release packaging now attests the native Python DSV4 encoder, nested affine defaults, mixed module quantization metadata, and the clean JANG runtime source, with separate Sequoia and Tahoe packaging/notarization contracts.

    7 details in Everything

11. v1.6.19 Jul 31, 2026 · issue 032

    vMLX 1.6.19 adds native-MTP diagnostics, model-owned sampling defaults, separate macOS Tahoe builds, and scoped sparse-cache blocks for MiniMax M3.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.6.19 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.6.19

    • Adds native-MTP health and profiling telemetry exposing bounded cache-lifecycle, acceptance, and phase-timing metrics for Qwen MTP models.
    • Introduces model-owned generation defaults: temperature, Top P, Top K, repetition penalty, and output limits now hydrate from the selected model bundle in new and restored chats.
    • Ollama requests now preserve an explicitly supplied top_k value, including 0, through request translation.
    • Automatic prompt limits are now derived from the actually loaded model's configuration, bounded by its declared context ceiling and current memory budget.
    • MiniMax M3 native sparse-cache blocks are scoped to the prompt-prefillprefillThe 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. shape, preventing numerically incorrect Lightning Indexer state restoration across mismatched prefill shapes.

    +7 more show less

    • Cache namespaces for looped-transformer bundles now include the effective repeated-layer layout, preventing cross-incompatible-loop-count cache reuse.
    • Compatible Nanbeige looped-transformer JANG bundles now validate loop count, effective cache-slot count, and runtime metadata before generation, rejecting inconsistent bundles.
    • MiniMax M2-family automatic cache storage defaults to the q8 correctness-first policy.
    • Persistent cache payload encoding is moved off the inference-completion path, reducing synchronous CPUCPUThe main general-purpose processor in a computer. It runs application instructions and coordinates work with memory and other hardware. work at end of request.
    • KaTeX rendering gains improved disambiguation of literal currency symbols from inline math and unclosed dollar delimiters.
    • New settings and compatibility messages are localized across English, Spanish, Japanese, Korean, and Chinese interfaces.
    • Separate Apple-silicon downloads are now provided for macOS Tahoe and Sequoia-compatible systems.

    12 details in Everything

12. v1.6.18 Jul 27, 2026 · issue 028

    vMLX 1.6.18 adds a process-wide SSD cache budget with LRU eviction and detailed cache reporting.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.6.18 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.6.18

    • Adds a process-wide SSD cache budget with coordinated LRU eviction and off-request-path publication for improved prefix reuse management.
    • Introduces detailed cache reporting covering attempted prefixes, applied reuse, uncached suffixes, reconstruction, fallback prefillprefillThe 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., and aggregate disk usage.
    • Supports MiniMax tool arguments emitted as request-schema XML children.
    • Retains media across Responses tool continuations and enables local attachment playback in the Electron app.

    4 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !This release requires JANG 2.5.34.
13. v1.6.16 Jul 23, 2026 · issue 027

    vMLX 1.6.16 ships KaTeX-backed Electron renderer for chat and reasoning surfaces and hardens bundled engine provenanceprovenanceEvidence recording where a software artifact came from and how it was built, including its source and build process. It helps a consumer check that the artifact matches a trusted build..

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.6.16 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.6.16

    • Adds KaTeX-backed Electron renderer path for chat and reasoning surfaces.
    • Hardens packaged engine provenance so signed app sessions use bundled Python/vmlx_engine instead of stale development or system virtualenvs.

    2 details in Everything

14. v1.6.12 Jul 20, 2026 · release notes

    vMLX 1.6.12 adds block-disk L2 as a usable prefix tier, hardens q4 TurboQuant KV storage, and fixes gateway lifecycle on client disconnect.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.6.12 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.6.12

    • Enables block-disk L2 as a usable prefix tier even when paged RAMRAMWorking memory that holds programs and data while a computer is running. Its contents usually disappear when power is removed. is disabled, while preserving the RAM-first then disk-refault hierarchy when both tiers are active — covering partial-prefix restore, eviction, immediate-stop durability, and restart refault.
    • Hardens q4 TurboQuant storage for eligible KV components and hybrid attention paths, keeping JANG affine, JANGTQ/MXTQ Hadamard-codebook, and base MLX MXFP as distinct formats; architecture-specific cache policies for DSV4, openPangu, MiniMax-M3, Gemma mixed-SWA, and other typed/native cache paths are preserved.
    • Extends media ownership and cache-key coverage to Step and Nemotron Omni image, video, and audio paths while keeping MiniMax-M2.7 as text-only.
    • Tightens progressive reasoning, content, and tool streaming plus terminal/usage ordering across Qwen, HY3, Step, MiniMax, Anthropic, Ollama, and Responses routes, including no-tool prompts and rejected/incomplete native tool markup.
    • Improves model-derived UI accuracy for quant labels, parser/reasoning/cache settings, single-model swaps, session PID/sleep lifecycle, eager Start materialization, port/LAN rollback, and missing-model repoint/removal.

    5 details in Everything

15. v1.6.11 Jul 19, 2026 · release notes

    vMLX 1.6.11 adds progressive streaming across Responses, Chat Completions, and Electron, plus extended L2 cache reuse for hybrid/mixed-attention model families.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.6.11 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.6.11

    • Streams post-reasoning and post-tool answers progressively across Responses, Chat Completions, and Electron UI instead of withholding output until completion.
    • Extends architecture-typed prefix, paged, and block-disk L2 cache reuse to hybrid and mixed-attention model families.
    • Adds media and cache routing support for Qwen, Bonsai, Gemma, Step, and MiniMax-M3 model families (MiniMax-M2.7 remains text-only).
    • Hardens multi-turn tool continuation, terminal events, cancellation and recovery, and reasoning/content separation.
    • Ships two separate macOS DMG builds: vMLX-1.6.11-sequoia-arm64.dmg (macosx_14_0_arm64, macOS 14.5+) and vMLX-1.6.11-tahoe-arm64.dmg (macosx_26_0_arm64).

    5 details in Everything

16. v1.6.10 Jul 15, 2026 · release notes

    vMLX 1.6.10 adds JANG affine 1-bit/ternary bundle loading, paged block-disk L2 cache on by default, and Responses API as the Electron chat default.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.6.10 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.6.10

    • Enables paged block-disk L2 cache by default in the Electron UI, with an explicit opt-out available.
    • Loads schema-2 JANG affine 1-bit and ternary text/VLMVLMA vision-language model that works with images and text together, such as answering questions about a picture or describing what it contains. bundles with per-module quantizationquantizationStoring 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. overrides.
    • Keeps local Electron chat on the OpenAI /v1/responses API by default, with chat sampler defaults derived from the model bundle configuration.
    • Preserves Qwen 3.5 hybrid GDN/SSM + attention cache routing, VL/video tensors, TurboQuant attention KV, and async SSM rederive.

    4 details in Everything

17. v1.6.7 Jul 13, 2026 · issue 036

    vMLX 1.6.7 enables paged KV cacheKV cacheMemory 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. by default for major text model families and turns reasoning ON by default for all reasoning-capable models.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.6.7 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.6.7

    └──▷ TRY IT
    Override the new paged-cache default to explicitly disable it for a model family that behaves unexpectedly with block-based caching.
    copy

    $ vmlx serve mlx-community/Qwen3-8B-4bit --no-paged-cache

    Explicitly opt in to paged KV cache on a model where you want to confirm block-based deduplication is active.
    copy

    $ vmlx serve mlx-community/Qwen3-8B-4bit --use-paged-cache

    • Paged KV cache now defaults ON for autodetected text families (Qwen3.5/3.6, Zaya, Nemotron-H, LFM2.5, Laguna, Hy3, Step-3.7) when continuous batchingcontinuous batchingA serving technique that adds and removes requests from the GPU batch at every step instead of waiting for a whole batch to finish. It raises throughput on shared inference servers. and prefix cache are active; --use-paged-cache / --no-paged-cache always override the default.
    • Reasoning mode now defaults ON for every reasoning-capable family; max_thinking_tokens is honored and gated on both streaming and non-streaming paths.

    2 details in Everything

18. v1.5.66 Jun 19, 2026 · release notes

    vMLX v1.5.66 adds MiniMax-M3 autodetection, native MSA prefix-cache handling, and Metal wired-memory guidance for Apple Silicon.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.66 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.66

    • Adds MiniMax-M3 autodetection with reasoning/tool parser wiring and model-owned generation defaults, including native MSA prefix-cache handling.
    • Surfaces model-owned max output behavior in both UI and CLI settings for MiniMax-M3 launch defaults.
    • Adds user-facing Metal wired-memory guidance when a model launch fails due to Insufficient Memory on Apple Silicon.
    • Retains Gemma 4 VL settings and media stress harness coverage for the scoped release path.

    4 details in Everything

19. v1.5.65 Jun 18, 2026 · release notes

    vMLX 1.5.65 adds MiniMax-M3 and Gemma 4 VL compatibility with reasoning mode, VL image input, and multi-API support.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.65 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.65

    • Adds MiniMax-M3 (MM3) model support with startup configuration: prefix cache enabled, paged cache disabled, disk prompt cache enabled, JITJITA runtime compilation strategy where code is translated from an intermediate form to native machine instructions at execution time, giving tools the speed of compiled code without sacrificing dynamic flexibility. disabled, native MSA/Lightning sparse cache storage, and generic TurboQuant KV disabled.
    • Supports MM3 reasoning mode in three states: off, on, and auto, across UI and API paths.
    • Supports visible content emission and exact tool calls for MM3 across UI and API paths.
    • Extends API compatibility to cover Responses, Chat Completions, Anthropic-compatible, and Ollama-compatible endpoints for MM3.
    • Adds streaming support for MM3 across all covered API paths.

    +3 more show less

    • Adds VL image input support for MM3.
    • Adds prefix-cache reuse telemetry for MM3.
    • Updates latest.json to point to the notarized 1.5.65 Sequoia and Tahoe DMGs.

    8 details in Everything

20. v1.5.64 Jun 18, 2026 · release notes

    vMLX 1.5.64 adds MiniMax-M3 and Gemma 4 VL model support with streaming reasoning fixes and cache configuration defaults.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.64 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.64

    • Adds MiniMax-M3 (MM3) model support across live UI, Chat Completions, Responses, Anthropic-compatible, Ollama-compatible, and streaming tool/image paths.
    • MiniMax-M3 startup defaults: prefix cache enabled, paged cache disabled, disk prompt cache enabled, JITJITA runtime compilation strategy where code is translated from an intermediate form to native machine instructions at execution time, giving tools the speed of compiled code without sacrificing dynamic flexibility. disabled, native MSA/Lightning sparse cache storage enabled, generic TurboQuant KV disabled.
    • Adds Gemma 4 E2B, E4B, 12B, 26B, and 31B JANG_4M VL variants with live UI, API, and cache coverage, plus 26B and 31B MXFP4MXFP4A 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. visual row support.
    • Streaming reasoning-only responses now produce a bounded visible-answer pass for supported reasoning models, preventing blank output when the model emits hidden reasoning only.
    • Updates latest.json to point to the notarized 1.5.64 Sequoia and Tahoe DMGs, which are Developer ID signed, Apple notarized, stapled, and Gatekeeper accepted.

    5 details in Everything

21. v1.5.63 Jun 18, 2026 · release notes

    vMLX v1.5.63 adds MiniMax-M3 and Gemma 4 E2B/E4B/12B MXFP4MXFP4A 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. vision/text compatibility with multimodalmultimodalAble 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. routing and reasoning controls.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.63 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.63

    • Adds MiniMax-M3 compatibility with multimodal routing, reasoning controls (off/on/auto), tool calls, and prefix-cache hit stability across Chat/Responses/Anthropic/Ollama endpoints.
    • Adds Gemma 4 E2B, E4B, and 12B MXFP4 vision/text compatibility with reasoning controls, image input, tool calls, and cache hits across Chat/Responses/Anthropic/Ollama and streaming surfaces.
    • Preserves model-owned generation defaults for supported startup paths.
    • MiniMax-M3 enforces its native MSA/Lightning sparse cache policy: paged cache disabled, TurboQuant KV skipped for native MSA, and JITJITA runtime compilation strategy where code is translated from an intermediate form to native machine instructions at execution time, giving tools the speed of compiled code without sacrificing dynamic flexibility. disabled for the dynamic sparse-cache path.

    4 details in Everything

22. v1.5.56 Jun 5, 2026 · release notes

    vMLX 1.5.56 preserves VMLX_VLM_IMAGE_PREFILL_BUFFER_GB overrides while switching to dynamic Metal working-set-based buffer sizing.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.56 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.56

    • Retains VMLX_VLM_IMAGE_PREFILL_BUFFER_GB as a hard override for users who need an explicit VLMVLMA vision-language model that works with images and text together, such as answering questions about a picture or describing what it contains. image prefillprefillThe 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. buffer guard.
    • Defaults Gemma 4 unified chat/image requests to visible-answer mode unless thinking is explicitly requested.

    2 details in Everything

23. v1.5.52 Jun 2, 2026 · release notes

    vMLX 1.5.52 reads greedy generation defaults (temperature, top-p, top-k) directly from JANG model metadata and ships Sequoia- and Tahoe-compatible DMGs.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.52 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.52

    • Reads and applies model-owned greedy generation defaults (temperature 0.00, top_p 1.0, top_k off) declared in JANG bundle metadata, so sampling behaviour is driven by the model file rather than hidden app-side assumptions.
    • Updates the local Step-3.7-Flash JANG metadata path so startup settings (default temperature=0.0, top_p=1.0, top_k=0) derive from model files; explicit request parameters still override model defaults.
    • Builds both public DMG flavors from the same source: Sequoia-compatible macosx_14 wheels and Tahoe-native macosx_26 wheels, with Developer ID signing, notarization, stapling, and Gatekeeper acceptance.

    3 details in Everything

24. v1.5.47 May 22, 2026 · release notes

    vMLX 1.5.47 adds Tahoe-native DMG for macOS Tahoe, tightens parser boundaries across model families, and expands DSV4 release guards.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.47 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.47

    • Adds a Tahoe-native DMG (vMLX-1.5.47-tahoe-arm64.dmg) for macOS Tahoe users who want native macosx_26 MLX wheels, alongside the existing Sequoia-compatible DMG (vMLX-1.5.47-sequoia-arm64.dmg).
    • Adds stronger parser boundaries so non-reasoning families such as Qwen2 and Qwen2-VL no longer silently inherit reasoning parsers, while MiniMax, DSV4, ZAYA, Hy3, Gemma, and GLM parser rows remain covered.
    • Adds DSV4-specific release guards for stale native-MTP launch args, cache/prefix policy, and JANG source parity.

    3 details in Everything

25. v1.5.45 May 20, 2026 · release notes

    vMLX 1.5.45 adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. config auto-discovery, chained tool calls across multi-round chats, and localized What's New notices.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.45 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.45

    └──▷ TRY IT
    Let vMLX auto-discover your MCP configuration without specifying a launch flag — just place your config in the working directory.
    copy

    $ # Place your MCP config in the project root, then launch normally
    # vMLX will auto-discover mcp.json or mcp.yaml without any extra flag
    echo '{"mcpServers": {"my-server": {"command": "npx", "args": ["-y", "my-mcp-server"]}}}' > mcp.json
    vmlx

    • MCP config auto-discovery now loads mcp.json or mcp.yaml from the working directory or user config paths without requiring a launch flag, and MCP policy flows through sessions and gateway routes.
    • Built-in coding tools and MCP tools can now chain across multi-round chats using structured tool_calls, with a text-chat watchdog that clears stuck Generating tool call states.
    • Adds support for Qwen3.6 native MTP, JANG/JANGTQ group-size detection, DSV4 cache policy, and TurboQuant/L2/paged cache telemetry with stricter release gates.
    • The in-app What's New notice is now localized in English, Spanish, Japanese, Korean, and Chinese, covering MCP, MTP, cache, and updater safety notes.

    4 details in Everything

26. v1.5.39 May 17, 2026 · release notes

    vMLX v1.5.39 adds native MTP health reporting with depth, scope, tensor counts, and gating reasons, plus Qwen 3.6 and Kimi K2.6 runtime support.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.39 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.39

    • Native MTP health/status now reports depth, scope, tensor counts, and gating reasons.
    • Supports Qwen 3.6 MTP artifacts defaulting to D3, with opt-in model-local tuning sidecars.
    • Qwen 3.6 MTP+VL app detection preserves multimodalmultimodalAble 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. behavior for artifact-backed affine-JANG MTP+VL bundles when indexed MTP and vision tensors are present.
    • Adds Source/PyPI runtime patch parity for Kimi K2.6 MLA and Gemma 4 vision pixel-value handling.

    4 details in Everything

27. v1.5.37 May 16, 2026 · release notes

    vMLX 1.5.37 adds max_prompt_tokens, max_context_tokens, and max_context aliases for prompt context admission across OpenAI, Anthropic, Ollama, and gateway paths.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.5.37 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.5.37

    └──▷ TRY IT
    Cap how much prompt context is admitted per request when calling through the OpenAI-compatible gateway, to stay within model token limits.
    copy

    $ max_prompt_tokens=4096 max_context_tokens=8192

    • Adds max_prompt_tokens, max_context_tokens, and max_context aliases for controlling request/session prompt context admission across OpenAI, Anthropic, Ollama, and gateway paths.
    • Applies model-declared generation defaults at session startup, removing panel-side sampling fallbacks.
    • Scopes sampling overrides to the current chat or API request instead of carrying stale per-model state across requests.

    3 details in Everything

28. v1.4.7 May 1, 2026 · release notes

    vMLX v1.4.7 promotes VMLX_OMNI_BACKEND to a CLI flag, unlocking a fast native MLX Stage-2 path for Nemotron-Omni.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.4.7 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.4.7

    └──▷ TRY IT
    Opt in to the Stage-2 native MLX backend for significantly faster Nemotron-Omni inference on M4 Max without setting an environment variable.
    copy

    $ vmlx --omni-backend stage2

    • Promotes VMLX_OMNI_BACKEND from an environment variable to a CLI flag, allowing the panel and Swift Settings UI to switch the Nemotron-Omni multimodalmultimodalAble 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. path between Stage-1 (bit-exact PyTorch bridge, default) and Stage-2 (native MLX RADIO + Parakeet, ~15–21x faster encoders and ~82 tok/s decode on M4 Max).
    • Adds Stage-2 native MLX backend for Nemotron-Omni, delivering ~15–21x faster encoders and ~82 tok/s decode on M4 Max hardware (off by default pending Wave-4 quality validation).
    • Confirms Qwen 3.6 VL is fully wired on both the Python and Swift sides.

    3 details in Everything

29. v1.4.2 May 1, 2026 · release notes

    vMLX v1.4.2 adds Laguna and Mistral-Medium-3.5 loaders, extends multimodalmultimodalAble 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. routing to /v1/responses and /v1/messages, and fixes Omni bundle boot.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.4.2 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.4.2

    └──▷ TRY IT
    Serve an Omni bundle that previously crashed on boot, then send an image via the OpenAI Responses endpoint using the newly supported input_image type.
    copy

    $ vmlx-serve <omni-bundle>

    • Extends OmniMultimodalDispatcher multimodal routing (RADIO + Parakeet encoders) to /v1/responses and /v1/messages, joining /v1/chat/completions — all 6 modality surfaces (text, image, audio, video) now dispatch correctly for Nemotron-3-Nano-Omni.
    • Adds _MEDIA_TYPES entries in _responses_input_to_messages for OpenAI Responses canonical types input_image, input_video, input_audio, and input_text, normalizing them to chat-completions envelopes.
    • Adds loaders/load_laguna.py routing through jang_tools.laguna.runtime.load for Laguna (poolside) 33B/3B agentic-coding MoEMoEA 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. (hybrid SWA+full attention, 256 routed experts top-8 + 1 shared, dual-RoPERoPERotary 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.); registered as a silver-tier model.
    • Adds loaders/load_mistral3.py for Mistral-Medium-3.5-128B (ministral3 inner), a dense GQA 96/8 model with 256K YaRNYaRNA method for extending a model's context window by adjusting rotary position embeddings. It helps a model handle sequences longer than those used during its original training. context and PIXTRAL vision; registered as a silver-tier model.

    4 details in Everything

30. v1.4.1-swift Apr 30, 2026 · release notes

    vMLX v1.4.1-swift adds MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. HTTP/SSESSEServer-Sent Events, a browser API and HTTP-based protocol where a server pushes a stream of text events to a client over a single long-lived connection, giving tools a lightweight alternative to WebSockets for one-way real-time updates. transport and NemotronHOmni VLMVLMA vision-language model that works with images and text together, such as answering questions about a picture or describing what it contains. support alongside UI and cache improvements.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.4.1-swift https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.4.1-swift

    • Adds MCP HTTP/SSE transport (real implementation) via mlxstudio#31, enabling Model Context ProtocolModel Context ProtocolModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. over HTTP and Server-Sent Events.
    • Adds NemotronHOmni VLM factory wiring and capability detector, bringing support for the NemotronHOmni vision-language modelvision-language modelA vision-language model that works with images and text together, such as answering questions about a picture or describing what it contains..
    • Adds Delete model files… context menu item in the chat picker for in-app model file management.

    3 details in Everything

31. v1.4.1 Apr 30, 2026 · release notes

    vMLX v1.4.1 adds model deletion API, image-gen cancellation, MCPMCPModel Context Protocol, a standard for connecting AI applications to external tools and data sources. A compatible server exposes capabilities an assistant can use through a common interface. HTTP+SSESSEServer-Sent Events, a browser API and HTTP-based protocol where a server pushes a stream of text events to a client over a single long-lived connection, giving tools a lightweight alternative to WebSockets for one-way real-time updates. transport, and vmlx[mxtq] install extras.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.4.1 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.4.1

    └──▷ TRY IT
    Install vMLX with MXTQ/JANG/DSV4/Nemotron-Omni bundle support in one step.
    copy

    $ pip install -U "vmlx[mxtq]==1.4.1"

    Abort a running image generation job server-side — useful when the client disconnects or the user clicks Stop.
    copy

    $ curl -X POST http://localhost:8080/v1/images/cancel

    • Adds POST /api/delete endpoint for model deletion with model-directory bounding (refuses paths outside the model root and path-traversal attempts).
    • Adds POST /v1/images/cancel endpoint and ImageGenEngine.cancel() method, plus a disconnect-watchdog so browser close or the Stop button truly aborts image generation instead of OOMOOMAn 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.-after-cancel.
    • Adds [mxtq] and [jang] install extras (pip install vmlx[mxtq]) for JANG/MXTQ/DSV4/MiniMax-JANG/Nemotron-Omni bundles.
    • Accepts Anthropic-style reasoning: {effort: ...} and reasoning: {type:'enabled', budget_tokens:N} request bodies, normalizing them into reasoning_effort / enable_thinking internally.
    • Replaces the stdio-only MCP guard with a full HTTP+SSE transport implementation on the Swift side: streaming via URLSession.bytes(for:), legacy endpoint-event handling, and Streamable-HTTP inline-JSON-RPCRPCRemote Procedure Call, a protocol that lets a program execute a function on another machine as if it were local, giving cyber tools a structured way to communicate across processes or hosts without managing raw sockets. fallback.

    +2 more show less

    • Adds Delete model files… context menu in the Swift app chat picker, backed by the new POST /api/delete endpoint.
    • Registers NemotronH-Omni in VLMModelFactory and the processor registry; auto-promotes nemotron_h → nemotron_h_omni when config_omni.json is present alongside config.json.

    7 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !macOS 14.4.x is no longer supported — the bundled MLX wheel requires libc++ symbols added in macOS 14.5 (Sonoma/Sequoia/Tahoe); running on 14.4.x will exit with an actionable error.
32. v1.3.34 Apr 9, 2026 · release notes

    vMLX v1.3.34 adds deferred SSM re-derive for hybrid thinking models, cutting multi-turn cache misses on Nemotron and Qwen3.5-VL.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.3.34 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.3.34

    • Adds deferred SSM re-derive: after generation completes on hybrid SSM + thinking models (Nemotron, Qwen3.5-VL), the scheduler queues an idle-time prefillprefillThe 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. pass on prompt tokens only, storing clean SSM state so the next conversation with the same prompt prefix gets KV + SSM cache hits instead of a full re-prefill.
    • Adds _ensure_mx_array() for dtype normalization: input_ids is cast to mx.int32 and pixel_values to mx.array, enabling batched VLMVLMA vision-language model that works with images and text together, such as answering questions about a picture or describing what it contains. inference on Mistral 3 / Pixtral.
    • mx.compile() now targets language_model.model only for VLM engines, preserving the wrapper's .config attribute and fixing JITJITA runtime compilation strategy where code is translated from an intermediate form to native machine instructions at execution time, giving tools the speed of compiled code without sacrificing dynamic flexibility. + MLLM compatibility.
    • _get_n_kv_heads() now scans cfg.text_config as a nested fallback, resolving 100% prefix-cache miss on Gemma 4 VLM where num_key_value_heads lives inside model.config.text_config.

    4 details in Everything

33. v1.3.31 Apr 9, 2026 · release notes

    vMLX v1.3.31 adds cross-session cache sharing, Aho-Corasick stop matching, Gemma 4 support, and a public /v1/cache stats endpoint.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.3.31 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.3.31

    └──▷ TRY IT
    Cap prefix cache memory consumption to avoid OOMOOMAn 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. on long-context workloads.
    copy

    $ vmlx serve --prefix-cache-max-bytes 4294967296 <model>

    • Adds --prefix-cache-max-bytes CLI flag to enforce a byte-budget eviction limit on the prefix cache.
    • Exposes size and max_entries fields on the /v1/cache stats endpoint for live cache introspection.
    • Adds cache_type column and per-type bucket eviction to the disk SQLite cache, with priority eviction and last-fetch type round-trip in L1 backfill.
    • Extracts SSMCompanionCache to vmlx_engine/utils/ssm_companion_cache.py with a per-entry is_complete flag for SSM companion cache lifecycle tracking.
    • Adds compute_model_cache_key() for content-derived cache identity that survives JITJITA runtime compilation strategy where code is translated from an intermediate form to native machine instructions at execution time, giving tools the speed of compiled code without sacrificing dynamic flexibility. reload and prevents cross-config pollution.

    +13 more show less

    • Adds _compute_segment_boundaries helper to engine/batched.py, firing segment boundaries on memory-aware and paged schedulers (not just legacy).
    • Adds is_hybrid_ssm_cache, is_hybrid_ssm_model, and is_hybrid_ssm_config polymorphic helpers for hybrid SSM model detection.
    • Adds is_mla_model centralized helper that walks top → text_config → language_config → llm_config for MLA model detection.
    • Adds Gemma 4 native text MoEMoEA 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. local-port shim with forward-compat that defers to upstream when available.
    • Sets trust_remote_code=True as the default in utils/tokenizer.py, unblocking custom tokenizers without manual configuration.
    • Enables LRU+Trie cross-session prefix sharing on MemoryAwarePrefixCache (production default) with system → user → assistant priority eviction.
    • Adds cache_type LRU and per-type bucket eviction on BlockAwarePrefixCache for paged and hybrid auto-switch scenarios.
    • Surfaces SSM companion cache stats (nbytes_mb) to the CachePanel UI.
    • Backports BatchMambaCache.lengths, advance, prepare, and finalize from mlx-lm 0.31.2.
    • Caches is_mllm_model() result in api/utils.py to eliminate repeated INFO log spam.
    • Adds scheduler.deep_reset wiring of reset_factory_cache() on shutdown.
    • Bumps mlx-lm dependency from 0.30.2 to 0.31.2.
    • Removes orphan disk-streaming artifacts (ssd_generate, streaming_wrapper, weight_index) superseded by the memory_cache and disk_cache architecture.

    18 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Orphan disk-streaming artifacts ssd_generate, streaming_wrapper, and weight_index have been removed; setups relying on them must migrate to the memory_cache + disk_cache architecture.
34. v1.3.25 Apr 3, 2026 · release notes

    vMLX v1.3.25 adds full Gemma 4 support with reasoning/tool parsers and a new --default-enable-thinking CLI flag.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.3.25 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.3.25

    └──▷ TRY IT
    Start the vMLX server with thinking mode enabled by default for all Gemma 4 reasoning requests.
    copy

    $ vmlx serve --default-enable-thinking

    • Adds --default-enable-thinking CLI flag to control thinking mode at server startup.
    • Adds Gemma 4 tool parser supporting the native <|tool_call>call:name{args}<tool_call|> format for function-callingfunction-callingA 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. workflows.
    • Adds Gemma 4 reasoning parser handling the <|channel>thought...<channel|> protocol with Auto/On/Off thinking modes.
    • Adds <turn|> as a stop token alongside <eos> for Gemma 4 generation.
    • Adds <|tool_call> to tool call marker detection for streaming responses.

    5 details in Everything

35. v1.3.6 Mar 21, 2026 · release notes

    vMLX v1.3.6 adds disk streaming for RAMRAMWorking memory that holds programs and data while a computer is running. Its contents usually disappear when power is removed.-oversize models, Mistral Small 4 (119B) support, and universal MoEMoEA 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. JANG gate dequant.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.3.6 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.3.6

    └──▷ USE IT
    Configure a Mistral Small 4 session in a vMLX config file to use the new model type with MoE routing.
    yaml copy

    model_type: mistral4

    • Adds --stream-from-disk flag to load models larger than available RAM by leveraging macOS SSD paging (~7.4 GB/s); automatically disables all caching, and exposes a virtual memory budget slider in the session settings UI.
    • Adds model_type: mistral4 config value to support Mistral Small 4 (119B) with MLA attention, 128-expert MoE, and YaRNYaRNA method for extending a model's context window by adjusting rotary position embeddings. It helps a model handle sequences longer than those used during its original training. interleaved RoPERoPERotary 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., routed through the mlx-lm text path and mistral3 VLMVLMA vision-language model that works with images and text together, such as answering questions about a picture or describing what it contains. wrapper.
    • Extends JANG gate dequantization to any model with n_routed_experts > 0 (Mistral 4, DeepSeek V3, etc.), previously limited to Nemotron only.

    3 details in Everything

36. v1.3.0 Mar 20, 2026 · release notes

    vMLX v1.3.0 adds Nemotron-H support with hybrid SSM caching and session status banners with smooth token streaming.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.3.0 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.3.0

    • Adds Nemotron-H JANG engine support with gate dequant (8→2 bit), fc1/fc2 rename, and MTP filter (42 GB GPUGPUA graphics processing unit, a processor that performs many calculations in parallel. AI workloads use GPUs to speed up model training and inference., 46 tok/s).
    • Extends hybrid SSM full cache — prefix, paged, and disk caching — to Qwen3.5-A3B and Nemotron-H models.
    • Adds session status banners showing loading, sleeping, and stopped states in chat UI.
    • Adds renderer-side typewriter animation via requestAnimationFrame (rAF) for smooth token streaming.

    4 details in Everything

    └──▷ BREAKING ON UPGRADE

    • !Think-completion seed injection has been removed entirely.
    • !The STREAM_THROTTLE_MS configuration, _is_vlm_config() function, and vestigial weights parameter have been removed.
37. v1.0.7 Mar 18, 2026 · release notes

    vMLX v1.0.7 adds full mflux model support, img2img variation generation, and an Iterate UI for prompt-based image refinement.

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.0.7 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.0.7

    • Adds support for Klein 4B/9B, Kontext, Fill, Qwen Image, and FIBO mflux models in the model picker with correct class dispatch.
    • Enables img2img variation generation for Generate models via source image and strength parameter.
    • Adds Iterate UI action on any generated image to create variations with modified prompts.
    • Persists settings — steps, size, guidance, strength, and negative prompt — across restarts.
    • Adds Code tab placeholder (marked Coming Soon) to the UI.

    5 details in Everything

38. v1.1.0 Mar 6, 2026 · release notes

    vMLX v1.1.0 adds speculative decodingspeculative decodingSpeeding 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., embedding modelembedding modelLists 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. preloading, and server-wide generation defaults

    └──▷ GET THIS VERSION

    copy

    $ git clone --branch v1.1.0 https://github.com/jjang-ai/vmlx.git
    # already have the repo? check out this version:
    $ git checkout v1.1.0

    • Adds /v1/embeddings endpoint support via a pre-loadable separate embeddingembeddingLists 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. model, enabling semantic searchsemantic searchSearching by meaning rather than requiring the same words. It commonly compares vector representations to find related content, even when the wording differs. and RAGRAGRetrieval-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. workflows without a separate server.
    • Adds speculative decoding configuration with draft model and num-draft-tokens controls for 20–90% generation speedup.
    • Adds server-wide default temperature and top-p sliders for generation defaults.
    • Adds a dismissible auto-update checker banner that notifies when a new version is available.

    4 details in Everything