PyPI

mlx-vlm

0.7.4Blaizzy/mlx-vlm ↗New releases in your daily digest, changelog included.

MLX-VLM is a package for inference and fine-tuning of Vision Language Models (VLMs) and Omni Models (VLMs with audio and video support) on your Mac using MLX.

Last release 6 days ago
28 Sep 2026
Ships fairly regularly
a new release about every 2 weeks
Nearly every release is documented
notes for 60 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
2 years old
86 releases · first in 2024

Release timeline

86 releases since 2024

202420252026

Releases

0.7.428 Sept 2026
Generate multiple images per prompt via num_images (ming_image) by @Lazarus-931 in #2353
What's Changed

Generate multiple images per prompt via num_images (ming_image) by @Lazarus-931 in #2353

Feat/z image num images by @Lazarus-931 in #2354

Fix MiMo-V2.6 audio requests failing across threads in the server by @Sofille65 in #2352

Batched num_images generation for qwen_image, flux2, ideogram4 by @Lazarus-931 in #2355

Fix Qwen3.5-arch left padding for batches by @lucasnewman in #2357

Feat/qwen3 5 moe text by @Lazarus-931 in #2340

gpt-oss: parse harmony channels into reasoning and content by @Lazarus-931 in #2329

Fix the packed width of empty 3-, 5- and 6-bit quantized KV caches by @pierre427 in #2342

Fix small top-p and very low temperature sampling by @lucasnewman in #2358

Preserve interleaved image order in shared prompt formatting by @Blaizzy in #2362

Fix Qwen3-VL position embeddings with a quantized pos_embed by @rcfa in #2304

Fix three tool-parser failures: boolean property schemas, untyped Qwen3-Coder objects, unclosed parameters by @pierre427 in #2344

Feat/lfm2 colbert by @Lazarus-931 in #2360

Cleanup/model test ownership by @Lazarus-931 in #2364

Refactor/gliner model scope by @Lazarus-931 in #2363

Cleanup/moondream qwen test ownership by @Lazarus-931 in #2366

tests: remove restored Mistral test class by @Lazarus-931 in #2371

docs: link corrected multimodal checkpoints by @Lazarus-931 in #2372

Strip newlines after the thinking close tag across stream chunks by @pierre427 in #2369

Use model-based sampling defaults for image generation by @lucasnewman in #2373

Make streamed tool-call content match the non-streamed response by @pierre427 in #2365

Feat/lfm2 encoder by @Lazarus-931 in #2359

Return CacheList from Florence-2 and Nemotron-Parse make_cache by @Lazarus-931 in #2381

Fix Gemma 4 audio encoder attending one frame too far into the past by @yaanfpv in #2385

Support/internvl3 5 by @Lazarus-931 in #2375

DeepSeek V4.1 vision and language model support by @Lazarus-931 in #2214

Bump version to 0.7.4 by @lucasnewman in #2388

New Contributors

@Sofille65 made their first contribution in #2352

@pierre427 made their first contribution in #2342

@rcfa made their first contribution in #2304

@yaanfpv made their first contribution in #2385

Full Changelog: v0.7.3...v0.7.4

One column per month.

0.7.324 Sept 2026

Hemmingway-1 support by @Lazarus-931 in #2330

What's Changed

Hemmingway-1 support by @Lazarus-931 in #2330
MiMo-V2.6-Flash-RL support by @Lazarus-931 in #2327
limite-1b-violetto support by @Lazarus-931 in #2331
Qwen-Image-2.1 fixes by @lucasnewman in #2332
Add kv caching for Qwen-Image-2.1 by @lucasnewman in #2333
Fix MiMo-V2.6 native MXFP4 loading by @Lazarus-931 in #2337
Providing MiMo-V2.6 multimodal support by @Lazarus-931 in #2338
Ming-Image-0.1-Design (text-to-image) support by @Lazarus-931 in #2334
Release generation batch arrays without waiting for cyclic GC by @aprylewu in #2328
Include batching support for Mimo v2.6 e by @Lazarus-931 in #2339
Add rampart support and unify pii model's shared components by @Lazarus-931 in #2341
Enable batched audio for supported processors by @Lazarus-931 in #2348
Enable Gemma-3n audio batching by @Lazarus-931 in #2350
Bump version to 0.7.3 by @lucasnewman in #2351

New Contributors

@aprylewu made their first contribution in #2328

Full Changelog: v0.7.2...v0.7.3

Open source →

0.7.221 Sept 2026

Downmix stereo audio before resampling in load_audio by @kiarina in #2258

What's Changed

Downmix stereo audio before resampling in load_audio by @kiarina in #2258
Fix APC prefill and restore memory planning by @Blaizzy in #2262
Move APC cache layout logic into cache classes by @Blaizzy in #2263
fix: correct LFM2-VL/LFM2.5-VL vision tower and preprocessing by @ykhrustalev in #2266
deepseek_v4: restore five weights dropped during checkpoint key normalization by @Lazarus-931 in #2277
Remove computer use and agents examples by @Blaizzy in #2278
Account for built-in APC cache layouts without fallback by @Blaizzy in #2269
Use inherited APC memory profiles with specialized cache overrides by @Blaizzy in #2272
moe_offload: recognise expert tensors in checkpoints without a leading key prefix by @Lazarus-931 in #2284
Bound unused cache graphs during single-request decoding by @Blaizzy in #2271
Clarify APC stored and restored token counters by @Blaizzy in #2270
Guard _extend_cache against caches that cannot batch by @Lazarus-931 in #2286
Add port of Meta's Sapiens2 dense prediction model(s) by @eklipse2k8 in #2283
Fix Qwen3 Omni chunked prefill with visual embeds by @lucasnewman in #2265
Fix combined image and video inputs for Qwen3-Omni by @lucasnewman in #2287
Fix empty assistant content for Anthropic tool calls by @JakeATX in #2291
Fix GLM5-Next quantized head sanitization by @JakeATX in #2290
Add Bonsai 2 27B model by @lucasnewman in #2293
Reduce test suite and drive model tests from JSON cases by @Blaizzy in #2276
lfm2_vl: inherit the text LFM2 language model to enable speculative decoding by @tugot17 in #2280
Discover local models by default and expose loaded status by @Blaizzy in #2294
qwen3_vl_moe: feed video pixels through the vision tower by @Lazarus-931 in #2299
Support current moondream2 checkpoints and moondream3 mlx quants by @Lazarus-931 in #2301
Add support the the LLaDA-Image model family by @lucasnewman in #2306
Automatically handle repo variants of supported image models by @lucasnewman in #2303
feat: SAM 3D Objects for MLX by @eklipse2k8 in #2282
deepseek_vl_v2: change rope_traditional to False by @Lazarus-931 in #2323
fix(server): return 400 when a chat/omni model is sent to /v1/audio/* by @Lazarus-931 in #2320
Add Mistral Large 3 (675B VLM) support by @Lazarus-931 in #2308
docs: explain the server's speculative draft counters by @ykhrustalev in #2324
Fix Chat and Responses conversation parity by @Blaizzy in #2260
Add Qwen-Image-2.1 (text-to-image + edit) support by @Lazarus-931 in #2321
Bump version to 0.7.2 by @lucasnewman in #2326

New Contributors

@kiarina made their first contribution in #2258
@ykhrustalev made their first contribution in #2266
@tugot17 made their first contribution in #2280

Full Changelog: v0.7.1...v0.7.2

Open source →

0.7.114 Sept 2026

Add standalone DINOv2 model with register tokens and mask support by @eklipse2k8 in #2184

What's Changed

Add standalone DINOv2 model with register tokens and mask support by @eklipse2k8 in #2184
Add SAM 3.1 multiplex video tracking port by @eklipse2k8 in #2185
Accelerate exact Qwen MXFP4 verification by @YuhuaBillChen in #2106
Add Spark-X2.5 (spark2_5 hybrid-attention text model) by @Lazarus-931 in #2188
Return OpenAI segments for NeMo-alignment STT models (Parakeet/Canary) by @Lazarus-931 in #2189
Run distributed SFT reductions on the CPU stream to avoid Metal GPU timeout by @Lazarus-931 in #2191
Fix MiniCPM5 function-call parsing by @Blaizzy in #2190
Check native tool calling when adding models by @Blaizzy in #2192
Reuse shared APC prefixes, bound memory, and expose live settings by @Blaizzy in #2182
fix(qwen3_5): transpose Qwen3.8 Conv3d patch embed to NDHWC by @Taderich73 in #2186
Add opt-in exact 8-bit Qwen verifier fusion by @tjansn in #2117
Add TTS support for MiniCPM-o by @lucasnewman in #2193
Add GLM-5 Next support by @Blaizzy in #2127
Add video understanding support for Mage-VL by @lucasnewman in #2202
Use an exact prefix for thought marker removal by @lucasnewman in #2203
Clean up redundant and trivial tests by @lucasnewman in #2205
Add support for PP-DocLayout V3 (MLX) by @HashNuke in #2204
Hy4 model support by @Lazarus-931 in #2082
Extract tool-calling into a dedicated tools/ package by @Lazarus-931 in #2208
fix(moe_offload): match nested Experts.switch_glu path (Gemma 4) by @typemild in #2213
Add support for Indic-OCR by @HashNuke in #2198
Strip box markers in Glm46VMoEProcessor output by @Lazarus-931 in #2225
Restore optimized MXFP4 and Q3 verification by @YuhuaBillChen in #2209
mage_vl: remap flattened OptiQ model.* keys in sanitize by @Lazarus-931 in #2247
Revert "Restore optimized MXFP4 and Q3 verification" by @Blaizzy in #2251
Fix Mllama cross-attention masks during chunked prefill by @Blaizzy in #2249
Optimize Gemma 4 KV-shared prefill by @Blaizzy in #2157
aya_vision: strip response markers from generated text by @Lazarus-931 in #2252
Bump version to 0.7.1 by @lucasnewman in #2253

New Contributors

@Taderich73 made their first contribution in #2186
@HashNuke made their first contribution in #2204
@typemild made their first contribution in #2213

Full Changelog: v0.7.0...v0.7.1

Open source →

0.7.07 Sept 2026

Optimize Qwen4 n-gram shard gathers by @pyros-projects in #2034

What's Changed

Optimize Qwen4 n-gram shard gathers by @pyros-projects in #2034
Fix native quantized KV in Qwen MTP verification by @YuhuaBillChen in #1986
optize_qwen4exp by @JJJYmmm in #2037
Support Apodex 1.1 checkpoints by @Lazarus-931 in #2019
Add MTP support for Qwen3.8-Flash-Next by @lucasnewman in #2040
Anthropic route: keep in-conversation system messages in place (fixes prefix-cache defeat) by @Lazarus-931 in #2052
fix(server): transcribe m4a and other non-encodable audio uploads by @Lazarus-931 in #2051
fix(rfdetr): honor --show-boxes instead of ignoring it by @Anai-Guo in #2024
fix(sam3): let --threshold fall through to the documented per-task defaults by @Anai-Guo in #2025
Fix positioned top-k sampling for speculative generation by @tjansn in #2029
Add continuous batching support for Qwen3.8-Flash-Next by @lucasnewman in #2056
fix(qwen3.8): support official FP8 conversion by @Thump604 in #2058
Add APC disk caching for Qwen3.8-Flash-Next by @lucasnewman in #2059
Fix streamed responses after tool calls by @lucasnewman in #2062
Fix tool call parsing with quotes and nested values by @lucasnewman in #2064
fix(attention): materialize a quantized KV state before hand-rolled attention (#2060) by @Anai-Guo in #2065
Fix nemotron_h prefill crash when inputs and inputs_embeds are both passed by @Lazarus-931 in #2057
Add pr contributor reminder by @Lazarus-931 in #2068
Redesign APC for dense and hybrid model caches by @Blaizzy in #1960
moe-offloading by @Lazarus-931 in #1813
fix(cache): trim ChunkedKVCache on valid length, not buffer width (#2031) by @Anai-Guo in #2036
fix(recurrent_gemma): initialize nn.Module in LanguageModel by @Anai-Guo in #2075
Bump version to 0.7.0rc0 by @Blaizzy in #2083
perf(qwen3.8): adapt MTP draft depth by @Thump604 in #2046
feat(qwen3.8): add external PLE storage by @Thump604 in #2045
handling issue #2073, cache search locally by @Lazarus-931 in #2076
lfm2_vl: load flattened OptiQ checkpoints (#2088) by @Lazarus-931 in #2092
fix(video): resolve frame sampling in one place (#2050) by @Lazarus-931 in #2100
Support float-quantized compressed-tensors checkpoints by @Lazarus-931 in #2102
diffusion_gemma: strip channel scaffolding from decoded output by @Lazarus-931 in #2101
fix: require compatible mlx-audio for VoiceChat by @Thump604 in #2110
fix: stop generation on the tokenizer's EOS when the config omits it by @Lazarus-931 in #2112
Add LLaVA-OneVision by @Lazarus-931 in #2115
Fix phantom KV in ragged batched speculative rollback by @Lazarus-931 in #2113
Add DeepSeek-V4 DSpark speculative drafter by @Lazarus-931 in #2097
Add LongCat-Flash-Lite-Sparse (LSA + n-gram) by @Lazarus-931 in #2063
muse_glimmer: load flattened OptiQ checkpoints (#2108) by @Lazarus-931 in #2118
fix(speculative): detect native MTP drafters via num_nextn_predict_layers by @Lazarus-931 in #2120
fix(generate): chunk prefill for prompts shorter than prefill_step_size by @Lazarus-931 in #2119
GLM-5.3 - 2026-08-28 by @Lazarus-931 in #2078
GLM-5.3 MTP - 2026-08-28 by @Lazarus-931 in #2079
Fix TurboQuant batch decode, MTP KV quantization, and attention sinks by @YuhuaBillChen in #1822
Add OmniParser v2 YOLO11 support by @Lazarus-931 in #2122
fix(florence2): load legacy large-ft checkpoints on transformers>=5.14 by @Lazarus-931 in #2123
Add Video Depth Anything (video_depth_anything) by @eklipse2k8 in #2104
Strip Muse Glimmer channel scaffolding from generated output by @Lazarus-931 in #2129
sam3: fix pred_boxes inflation and offset (RoPE scale, RPB grid, ref_point_head) by @Lazarus-931 in #2128
Add OpenAI Privacy Filter model support by @lucasnewman in #2131
Support multiple tool call invocations at once by @lucasnewman in #2132
Require MLX 0.32.2 for faster speculative decoding by @Blaizzy in #2133
Add DSpark support for Gemma 4 by @Lazarus-931 in #2137
qwen3_vl: render per-group video timestamps like the reference processor by @Micha0827 in #2142
Add DeepSeek-V4 Flash Vision EXP support by @byoan in #2136
granite4_vision: align deepstack injection with chunked prefill via c… by @Lazarus-931 in #2143
Register the qwen3_vl_embedding auto-processor patch by @Lazarus-931 in #2146
Optimize exact Qwen verifier QMV unrolling by @Blaizzy in #2140
Keep quantized APC checkpoints packed by @Blaizzy in #2090
Normalize tokenizer padding in BatchGenerator by @Blaizzy in #1951
Add Nemotron-H speculative decoding support by @Lazarus-931 in #2145
Enable speculative decoding for Laguna by @Lazarus-931 in #2151
Fix ArraysCache trim crash on multi-turn reuse for hybrid models by @Lazarus-931 in #2152
mla: use the absorbed path for multi-token steps by @stooit in #2150
Fix Gemma4DSparkAttention AttributeError on first forward by @Lazarus-931 in #2158
Porting model K2-Horizon by @Lazarus-931 in #2153
Add Ornith 1.5 (Qwen3.5 VL) support and verify qwen3_5_moe MTP by @Lazarus-931 in #2125
Optimize Qwen4 QSA cache reuse and batched parity by @Blaizzy in #2126
Fix GLM and Kimi output control tokens by @Lazarus-931 in #2170
Add MoGe-3 monocular geometry estimation model by @eklipse2k8 in #2171
Add Z1T-0 (Extropic attention-free sparse LM) by @Lazarus-931 in #2166
Fix Unlimited-OCR prompt suffix by @Lazarus-931 in #2169
fix(aya_vision): reference patch_embeddings in the spatial_shapes branch by @Anai-Guo in #2173
Freeze base model when resuming LoRA training via --adapter-path by @Lazarus-931 in #2181
fix: declare jinja2 as a runtime dependency by @Shouta0108 in #2176
Bump version to 0.7.0 by @lucasnewman in #2175

New Contributors

@pyros-projects made their first contribution in #2034
@YuhuaBillChen made their first contribution in #1986
@tjansn made their first contribution in #2029
@eklipse2k8 made their first contribution in #2104
@Micha0827 made their first contribution in #2142
@stooit made their first contribution in #2150
@Shouta0108 made their first contribution in #2176

Full Changelog: v0.6.17...v0.7.0

Open source →

0.7.0rc031 Aug 2026pre-release

Optimize Qwen4 n-gram shard gathers by @pyros-projects in #2034

What's Changed

Optimize Qwen4 n-gram shard gathers by @pyros-projects in #2034
Fix native quantized KV in Qwen MTP verification by @YuhuaBillChen in #1986
optize_qwen4exp by @JJJYmmm in #2037
Support Apodex 1.1 checkpoints by @Lazarus-931 in #2019
Add MTP support for Qwen3.8-Flash-Next by @lucasnewman in #2040
Anthropic route: keep in-conversation system messages in place (fixes prefix-cache defeat) by @Lazarus-931 in #2052
fix(server): transcribe m4a and other non-encodable audio uploads by @Lazarus-931 in #2051
fix(rfdetr): honor --show-boxes instead of ignoring it by @Anai-Guo in #2024
fix(sam3): let --threshold fall through to the documented per-task defaults by @Anai-Guo in #2025
Fix positioned top-k sampling for speculative generation by @tjansn in #2029
Add continuous batching support for Qwen3.8-Flash-Next by @lucasnewman in #2056
fix(qwen3.8): support official FP8 conversion by @Thump604 in #2058
Add APC disk caching for Qwen3.8-Flash-Next by @lucasnewman in #2059
Fix streamed responses after tool calls by @lucasnewman in #2062
Fix tool call parsing with quotes and nested values by @lucasnewman in #2064
fix(attention): materialize a quantized KV state before hand-rolled attention (#2060) by @Anai-Guo in #2065
Fix nemotron_h prefill crash when inputs and inputs_embeds are both passed by @Lazarus-931 in #2057
Add pr contributor reminder by @Lazarus-931 in #2068
Redesign APC for dense and hybrid model caches by @Blaizzy in #1960
moe-offloading by @Lazarus-931 in #1813
fix(cache): trim ChunkedKVCache on valid length, not buffer width (#2031) by @Anai-Guo in #2036
fix(recurrent_gemma): initialize nn.Module in LanguageModel by @Anai-Guo in #2075
Bump version to 0.7.0rc0 by @Blaizzy in #2083
perf(qwen3.8): adapt MTP draft depth by @Thump604 in #2046
feat(qwen3.8): add external PLE storage by @Thump604 in #2045
handling issue #2073, cache search locally by @Lazarus-931 in #2076
lfm2_vl: load flattened OptiQ checkpoints (#2088) by @Lazarus-931 in #2092
fix(video): resolve frame sampling in one place (#2050) by @Lazarus-931 in #2100
Support float-quantized compressed-tensors checkpoints by @Lazarus-931 in #2102
diffusion_gemma: strip channel scaffolding from decoded output by @Lazarus-931 in #2101
fix: require compatible mlx-audio for VoiceChat by @Thump604 in #2110
fix: stop generation on the tokenizer's EOS when the config omits it by @Lazarus-931 in #2112
Add LLaVA-OneVision by @Lazarus-931 in #2115
Fix phantom KV in ragged batched speculative rollback by @Lazarus-931 in #2113
Add DeepSeek-V4 DSpark speculative drafter by @Lazarus-931 in #2097
Add LongCat-Flash-Lite-Sparse (LSA + n-gram) by @Lazarus-931 in #2063
muse_glimmer: load flattened OptiQ checkpoints (#2108) by @Lazarus-931 in #2118
fix(speculative): detect native MTP drafters via num_nextn_predict_layers by @Lazarus-931 in #2120
fix(generate): chunk prefill for prompts shorter than prefill_step_size by @Lazarus-931 in #2119
GLM-5.3 - 2026-08-28 by @Lazarus-931 in #2078

New Contributors

@pyros-projects made their first contribution in #2034
@YuhuaBillChen made their first contribution in #1986
@tjansn made their first contribution in #2029

Full Changelog: v0.6.17...v0.7.0rc0

Open source →

0.6.1726 Aug 2026

Fix MiniMax-M3 sparse-index compile re-trace leak (Metal 499k resource limit) by @ziomancer in #2016

What's Changed

Fix MiniMax-M3 sparse-index compile re-trace leak (Metal 499k resource limit) by @ziomancer in #2016
GLM-5.3-Flash support by @Lazarus-931 in #2030
Add Qwen3.8-Flash-Next support by @lucasnewman in #2032
Fix sectioned MRoPE split under mlx 0.32.2 by @Lazarus-931 in #2021
Add initial GLiNER2.5 support by @Lazarus-931 in #2020
Bump version to 0.6.17 by @lucasnewman in #2035

New Contributors

@ziomancer made their first contribution in #2016

Full Changelog: v0.6.16...v0.6.17

Open source →

0.6.1624 Aug 2026

Warn when resize_shape is ignored by a custom image processor by @eptan in #1829

What's Changed

Warn when resize_shape is ignored by a custom image processor by @eptan in #1829
Add Qwen3.8-27B support by @Lazarus-931 in #1899
fix(server): stop skipping KV cache quantization on "qat" model paths by @Lazarus-931 in #1955
Add num_image_tokens estimation for dynamic-resolution processors by @eptan in #1950
Use case-insensitive comparison for chat template detection by @lucasnewman in #1961
Refactor MTP split pathways by @Lazarus-931 in #1935
Remove surplus Qwen3-VL image placeholders by @byoan in #1941
Load a custom model class from a model_file declared in config.json by @noahzelezny in #1926
feat(server): alias reasoning_effort to reasoning_strength (Glimmer; supersedes #1843) by @nazerim in #1884
GLM-4.7 attention with TurboQuant APC support by @lucasnewman in #1975
fix(qwen3_omni): convert audio_feature_lengths to mel frames (#1620) by @Anai-Guo in #1948
Assert sanitize() idempotency in tests and document the invariant by @danmackinlay in #1821
Add Z-Image generation and conversion support by @scaryrawr in #1952
Add continuous batching support to PoolingCache by @lucasnewman in #1974
Fix instances of mx.arrays being implicitly used as integers by @lucasnewman in #1982
Add LFM2.5 DSpark speculative decoding by @Blaizzy in #1980
Make LFM2 DSpark exact: 2.90x dense and 1.24x MoE by @Blaizzy in #1987
State the cached-decode invariant once instead of per test by @Lazarus-931 in #1992
Stop CI from installing an older mlx than requirements allow by @Lazarus-931 in #1990
Tweaks to add-model skill to reduce review burden by @lucasnewman in #1973
Fix ArraysCache per-token Metal buffer leak on long hybrid-model generation (#1972) by @Lazarus-931 in #1984
Fix Qwen3-Omni decode assigning RoPE position 0 to every token by @Lazarus-931 in #1993
Fix DeepSeek-V4 quantization path aliases by @byoan in #1977
Expose --top-p, --top-k and --min-p on the generate CLI by @Lazarus-931 in #1994
Use integer Metal grid dimensions by @Lazarus-931 in #2006
fix(server): enforce tool_choice on the Anthropic /v1/messages route (#2000) by @Anai-Guo in #2002
Add Mage-Flow quantization support by @scaryrawr in #1979
Ensure APC is reset before unloading a model to avoid stale cache state persisting by @lucasnewman in #2010
Add exact Qwen3.8 27B DSpark speculative decoding by @Blaizzy in #1998
Bump version to 0.6.16 by @Blaizzy in #2012
Load text-only checkpoints that still advertise a vision tower by @Lazarus-931 in #1999
feat: ship a PEP 561 py.typed marker so mlx-vlm's annotations are usable (#1970) by @Anai-Guo in #1985
Add native ERNIE-Image generation and img2img support by @scaryrawr in #1954
Add Qwen3.8 27B DFlash2 speculative decoding by @Blaizzy in #2014

New Contributors

@eptan made their first contribution in #1829
@noahzelezny made their first contribution in #1926
@danmackinlay made their first contribution in #1821

Full Changelog: v0.6.15...v0.6.16

Open source →

0.6.1518 Aug 2026

Stop a batched row's answer depending on its neighbour's length by @Lazarus-931 in #1946

What's Changed

Stop a batched row's answer depending on its neighbour's length by @Lazarus-931 in #1946
Fix issues + tests with mlx 0.32.1 by @lucasnewman in #1949

Full Changelog: v0.6.14...v0.6.15

Open source →

0.6.1417 Aug 2026

Fix multimodal server prefill alignment by @Blaizzy in #1872

What's Changed

Fix multimodal server prefill alignment by @Blaizzy in #1872
Load pre-converted MLX Inkling checkpoints by @Lazarus-931 in #1880
Skip conv transposes for pre-converted SAM 3.1 checkpoints (#1871) by @Lazarus-931 in #1877
Add text and multimodal reranking endpoint by @Lazarus-931 in #1853
Keep pending padding metadata aligned after filtering by @byoan in #1890
Add PLaMo 2.1 VL support by @mitmul in #949
Fix Idefics3/SmolVLM LoRA training crashes (#1830) by @Lazarus-931 in #1852
Allow CacheList extraction with plain KVCache children by @byoan in #1870
Add Nemotron-VoiceChat-11B full-duplex speech model by @lucasnewman in #1817
Porting twenty four MLX-LM models by @Lazarus-931 in #1891
Add native sequence classifier rerankers by @Lazarus-931 in #1893
Add Nemotron-Parse (C-RADIO encoder + mBART decoder) model support by @axiom-of-choice in #1866
Fix thinking_budget silently ignored for self-opening thinking models by @Lazarus-931 in #1912
Fix SmolVLM inference with flattened image features by @Lazarus-931 in #1905
Stop a short first prompt from disabling prefix caching by @Lazarus-931 in #1901
Load text-only checkpoints of vision-language models by @Lazarus-931 in #1879
Support per-generator prefill step size overrides by @byoan in #1895
Fix zero-length rotating cache merge by @Lazarus-931 in #1916
Add GOT-OCR 2.0 model support by @axiom-of-choice in #1908
lfm2_vl: implement official image splitting (tiling) by @ivanfioravanti in #1885
fix(rope): support per-batch array offsets in eager RoPE by @nazerim in #1883
fix(molmo_point): honor point_color in draw_points_on_image by @Anai-Guo in #1887
Fix SmolVLM split-image token expansion by @atirna in #1922
feat(server): live runtime settings via /v1/settings (reload-on-change, no restart) by @Lazarus-931 in #1876
fix(models): honor a caller-supplied attention mask in five language models by @Anai-Guo in #1913
Use correct RoPE scaling for several models by @lucasnewman in #1927
Add Ministral3 embedder: reuse backbone with bidirectional mask by @Aryagarg23 in #1851
Add video subsampling to images fallback in the server by @lucasnewman in #1924
Fix PoolingCache APC restoration by @byoan in #1929
Fix EOS token for Idefics3 by @lucasnewman in #1936

New Contributors

@mitmul made their first contribution in #949
@axiom-of-choice made their first contribution in #1866
@nazerim made their first contribution in #1883
@Anai-Guo made their first contribution in #1887
@atirna made their first contribution in #1922

Full Changelog: v0.6.13...v0.6.14

Open source →

0.6.1312 Aug 2026

Preserve APC prefix reuse for growing prepared prompts by @byoan in #1713

What's Changed

Preserve APC prefix reuse for growing prepared prompts by @byoan in #1713
Flush the stream buffer on the last token by @Aryagarg23 in #1845
fix: tolerate null vision_config in quantization skip_vision path by @garnetlyx in #1777
Pin MLX_ENABLE_TF32=0 for the tests by @Aryagarg23 in #1846
Eager-eval cached RoPE arrays in the remaining variants by @Aryagarg23 in #1854
Fix LFM2.5-VL 3B checkpoint and image loading by @Blaizzy in #1863
Use model reasoning config and response templates by @Blaizzy in #1848
Bump version to 0.6.13 by @Blaizzy in #1868
Fix null content in assistant tool calls by @Lazarus-931 in #1864
Add Muse Glimmer DFlash assistant support by @Blaizzy in #1842
Add Cohere North Micro Vision support by @Blaizzy in #1804

New Contributors

@byoan made their first contribution in #1713
@Aryagarg23 made their first contribution in #1845
@garnetlyx made their first contribution in #1777

Full Changelog: v0.6.12...v0.6.13

Open source →

0.6.1210 Aug 2026

Fix LFM tool call parsing and bump version to 0.6.11 by @Blaizzy in #1792

What's Changed

Fix LFM tool call parsing and bump version to 0.6.11 by @Blaizzy in #1792
Fix Qwen3.5 ragged decode on GPUs that reject a 1024-thread threadgroup by @neil0306 in #1799
fix(quant): honor per-layer quantization overrides for wrapped text models by @Lazarus-931 in #1791
Fix Qwen3.5 structured output on checkpoints with a padded vocab by @Lazarus-931 in #1805
Add qwen3 text backbone support to llmjpvl (Jagle-VL) by @takuyaomi in #1801
feat(speculative): validate Laguna DFlash contract by @Thump604 in #1753
server: add --max-num-seqs to bound concurrent batch size by @Lazarus-931 in #1782
Honor explicit trust_remote_code in processor from_pretrained by @camparker in #1731
Fix /v1/responses image reference output by @lucasnewman in #1806
test: cover Chandra OCR2 config on the qwen3_5 path (#39) by @Lazarus-931 in #1809
Split unterminated reasoning correctly when the prompt opens the thinking block by @Lazarus-931 in #1811
embedding_loader: quantize pre-quantized weights before loading by @Lazarus-931 in #1773
Add support for Minimax-H3 by @lucasnewman in #1808
Keep serving when an optional model preload fails by @Lazarus-931 in #1814
Quick fix to properly recoginze Bonsai Image models. by @dcartman in #1816
Configure KV cache quantization per tensor by @Lazarus-931 in #1807
Add Muse Glimmer model support by @Blaizzy in #1838
Add Muse Glimmer ATEM tool parser by @Blaizzy in #1841

New Contributors

@camparker made their first contribution in #1731
@dcartman made their first contribution in #1816

Full Changelog: v0.6.10...v0.6.12

Open source →

0.6.104 Aug 2026

Add and improve MLX-VLM agent skills by @Lazarus-931 in #1747

What's Changed

Add and improve MLX-VLM agent skills by @Lazarus-931 in #1747
Skip full-sequence LM head during Gemma 4 prefill by @Lazarus-931 in #1750
Support current LFM2 feed-forward configs by @Blaizzy in #1786
Load Kimi K3 processor and tokenizer without remote code by @Blaizzy in #1787
Bump version to 0.6.10 by @Blaizzy in #1789
Convert Kimi K3 tiktoken tokenizers at runtime by @Blaizzy in #1790
Expose per-request speculative decoding acceptance stats in response timings by @Lazarus-931 in #1788

Full Changelog: v0.6.9...v0.6.10

Open source →

0.6.93 Aug 2026

Add Kimi K3 by @kernelpool in #1746

What's Changed

Add Kimi K3 by @kernelpool in #1746
Fix BPE detokenizer UTF-8 flush crash by @Blaizzy in #1705
Guard APC media token lookup for text-only models by @Blaizzy in #1733
fix(laguna): preserve provider tokenization contract by @Thump604 in #1719
Preserve audio in Gemma 4 video prompts by @Blaizzy in #1735
Fix double transpose of Nemotron Omni audio conv weights on MLX checkpoints by @takuyaomi in #1743
Fix prompt-cache prefix reuse corrupting rotating KV caches by @Lazarus-931 in #1748
Add Mage-VL (microsoft/Mage-VL, 4B codec-native VLM) by @xocialize in #1745
Porting embeddings by @Lazarus-931 in #1752
Add image input support to /v1/embeddings endpoint by @Lazarus-931 in #1762
Inkling Small by @pcuenca in #1756
Fix Inkling batching and caches with MTP by @lucasnewman in #1763
fix: isolate Qwen MTP shards during sanitization by @Thump604 in #1754
Inkling: decode fast path (fused sconv and MoE kernels, stacked qkvr, window-bounded attention) by @davidtorcivia in #1759
Add audio feature extractor for Inkling by @lucasnewman in #1767
Video input for models without a video_processor: frames fallback + Inkling native temporal pairs by @davidtorcivia in #1758
dflash: clear MLX cache and release verify capture each round by @Lazarus-931 in #1780
molmo: fix image-feature scatter and prompt formatting (unusable output) by @Lazarus-931 in #1783
Bump version to 0.6.9 by @Blaizzy in #1784

New Contributors

@kernelpool made their first contribution in #1746
@xocialize made their first contribution in #1745
@davidtorcivia made their first contribution in #1759

Full Changelog: v0.6.8...v0.6.9

Open source →

0.6.827 Jul 2026

Vendor standard models by @Lazarus-931 in #1702

What's Changed

Vendor standard models by @Lazarus-931 in #1702
Fix Laguna XS NVFP4 sanitizer by @Blaizzy in #1667
Add Mage-Flow image model by @lucasnewman in #1708
Authenticate inference API routes by @Blaizzy in #1714
Awq quant by @Lazarus-931 in #1666
Support mixed-precision compressed-tensors checkpoints (fp8 + NVFP4) by @Lazarus-931 in #1711
Fix Responses instructions for Qwen chat templates by @Blaizzy in #1716
Return model loading failures as bad requests by @Blaizzy in #1717
Allow image editing with FLUX.2-klein-4b by @lucasnewman in #1729
gpt_oss: honor per-layer mixed quantization (mxfp4 experts + 8-bit affine) by @Lazarus-931 in #1728
Support quantized FLUX.2 models by @lucasnewman in #1730
Add PLaMo 2.1 VL (plamo2vl) model support by @takuyaomi in #1739
Add llm-jp-4-vl (llmjpvl) model support by @takuyaomi in #1738
Fix chunked MRoPE positions across Qwen VL models by @Blaizzy in #1741
Fix multi-image SFT grid collation by @Blaizzy in #1744

New Contributors

@takuyaomi made their first contribution in #1739

Full Changelog: v0.6.7...v0.6.8

Open source →

0.6.723 Jul 2026

Support Laguna S by @Blaizzy in #1650

What's Changed

Support Laguna S by @Blaizzy in #1650
Add Inkling from Thinky support by @Lazarus-931 in #1637
Improve generation API type hints by @Blaizzy in #1599
fix(qwen3_omni_moe): treat audio feature mask as sample-domain (mel-frame lengths) by @Lazarus-931 in #1657
implement apc as a pluggable, capability driven cache adapter for #1629 by @Lazarus-931 in #1638
Fix compile validation by @Lazarus-931 in #1655
Add top-nσ sampler by @Lazarus-931 in #1652
chore(deps): make training-only datasets an optional [train] extra by @Lazarus-931 in #1659
add glm4_moe_lite_mtp drafter (GLM-4.7-Flash MTP split tool) by @samithaj in #1570
Bump version to 0.6.7 by @Blaizzy in #1661
fix(qwen3_omni_moe): accept batched (1, mel, T) audio features in encoder by @Lazarus-931 in #1656
feat(sampler): add p-less sampling by @Lazarus-931 in #1653
docs: document the ui extra required by the Gradio chat UI by @natejswenson in #1646
fix(lfm2_vl): make the training forward pass runnable for LoRA/SFT by @chintan-projects in #1631
Fix Qwen3 Omni deepstack: visual embeds were computed but never injected by @hellopahe in #1635
fix(qwen3-omni): keep mRoPE positions correct across prefill and decode by @Lazarus-931 in #1621
Add decoder_input_ids warm start for DiffusionGemma by @Blaizzy in #1589
Fix Laguna NVFP4 sanitize by @Blaizzy in #1664
Add typical-p (locally typical) sampler by @Lazarus-931 in #1663

New Contributors

@samithaj made their first contribution in #1570
@natejswenson made their first contribution in #1646
@chintan-projects made their first contribution in #1631

Full Changelog: v0.6.6...v0.6.7

Open source →

0.6.620 Jul 2026

vendor Gemma/Phi/GLM/Cohere text models and reuse shared MLPs by @Lazarus-931 in #1616

What's Changed

vendor Gemma/Phi/GLM/Cohere text models and reuse shared MLPs by @Lazarus-931 in #1616
fix(apc): exact hybrid + TurboQuant warm match live kv layout by @weklund in #1596
vendor 24 more text models as native packages by @Lazarus-931 in #1625
Speed up llguidance structured decoding on Qwen 3.5 by @Blaizzy in #1628
Improve server generation logging by @Blaizzy in #1634
Fix LFM2-VL tokenizer loading without PyTorch by @Blaizzy in #1640
Fix streaming reasoning classification by @Blaizzy in #1641
Fix streaming reasoning protocol compatibility by @Blaizzy in #1644
Bump version to 0.6.6 by @Blaizzy in #1645
Fix Qwen3-VL PIL video inputs by @Blaizzy in #1642

Full Changelog: v0.6.5...v0.6.6

Open source →

0.6.516 Jul 2026

Add deepseek_v3 and deepseek_v32 model types by @Lazarus-931 in #1517

What's Changed

Add deepseek_v3 and deepseek_v32 model types by @Lazarus-931 in #1517
port LFM2 into a shared components module and fix per-layer quantization overrides by @Lazarus-931 in #1513
Fix MRoPE decode positions in plain generate for Qwen VLMs by @Lazarus-931 in #1527
Fix Gemma4 video processor and audio weight loading by @Blaizzy in #1523
Bump version to v0.6.5 by @Blaizzy in #1529
Add end-to-end APC exact-mode correctness tests for hybrid models by @weklund in #1519
Add nemotron_h model type and remove mlx-lm from nemotron_h_nano_omni by @Lazarus-931 in #1518
Skip norm weight shift when loading MLX-format Qwen3.5 checkpoints by @Lazarus-931 in #1528
add qwen3 model type by @Lazarus-931 in #1531
add llama model type by @Lazarus-931 in #1530
add gemma3_text model type reusing gemma3 language internals by @Lazarus-931 in #1536
add mistral model type by @Lazarus-931 in #1535
Add structural tests for 8 untested model types by @weklund in #1539
Fix PaddleOCR-VL sanitize double-prefixing MLX-format checkpoints by @Lazarus-931 in #1540
Fix double-sanitize issue with Qwen3 Omni MoE by @lucasnewman in #1541
Optimize Qwen3 Omni audio hidden-state capture by @sxy-trans-n in #1522
fix(qwen3_5_mtp): guard norm +1.0 shift to raw-HF layout only by @Lazarus-931 in #1554
Enable APC with KV-cache quantization (single-request) by @weklund in #1534
feat(apc): adapter boundary for batch/SWA kv-bits (#1559) by @weklund in #1560
docs: APC works with KV-cache quantization by @weklund in #1565
Move Qwen3 Omni MRoPE to request-owned kwargs by @hellopahe in #1557
sam3: implement box/geometry prompt support in the detector by @Lazarus-931 in #1563
feat(apc): APC_TRACE + non-fatal layout self-check by @weklund in #1566
fix: multi-row continuous batching with KV quant (#1567) by @weklund in #1568
fix(qwen3_5): skip fully-padded rows in chunked batch prefill (#1558) by @Lazarus-931 in #1571
Fix Gemma 4 dtype leak by @Blaizzy in #1585
Guard Qwen3.5 norm weight shifts by @Blaizzy in #1556
Fix AutoProcessor.register string-first calls (transformers 5.13 issue) by @Lazarus-931 in #1581
Fix Qwen3.5 gated delta training async eval by @Blaizzy in #1586
fix(sam3): make sanitize idempotent for pre-converted MLX conv weights by @Lazarus-931 in #1550
Fix generation batch logits processor alignment by @Blaizzy in #1583
minimax_m3_vl: add pack_shared_expert config override for mixed-precision quantization by @avlp12 in #1544
Honor quantized_kv_start on the batch TurboQuant path by @Lazarus-931 in #1582
Add 1-bit inference kernel (Python-hosted) by @Blaizzy in #1597
vendor gated_delta kernel/ops from mlx-lm into qwen3_5 by @Lazarus-931 in #1593
vendor quantize/dequantize_model into mlx_vlm.quant_utils by @Lazarus-931 in #1594
Optimize 1-bit decode and fused greedy routing by @Blaizzy in #1598
Avoid thinking budget decode synchronization by @Blaizzy in #1601
Vendor qwen & deepseek text models by @Lazarus-931 in #1606
trainer: load text-model LoRA adapters via mlx-vlm's own loader by @Lazarus-931 in #1605
share MLP implementations getting refractored by @Lazarus-931 in #1607
share the vision multi-head Attention used 6 times by @Lazarus-931 in #1608
Preserve MLX attention masks during preprocessing by @Blaizzy in #1610
Add Chat Completions tool choice support by @Blaizzy in #1611
Allow transformers 5.14 and newer by @Blaizzy in #1612
Fix blocking continuous-batch prompt prefill by @Blaizzy in #1613

New Contributors

@weklund made their first contribution in #1519
@sxy-trans-n made their first contribution in #1522
@avlp12 made their first contribution in #1544

Full Changelog: v0.6.4...v0.6.5

Open source →

0.6.46 Jul 2026

Add support for TTS and STT endpoints in the server by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1358

What's Changed

Add support for TTS and STT endpoints in the server by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1358
Add prefill policy support for diffusion models by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1357
Fix diffusion mxfp4 embedding dequantization by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1356
fix: guard cache_offset == 0 against array truth-value error in continuous batching by @imi4u36d in https://github.com/Blaizzy/mlx-vlm/pull/1354
Fix DiffusionGemma generation from the server by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1368
Improve diffusion unmasking visualization by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1361
[Cohere] Fix streaming thinking parsing after prompt-opened by @Terrencezzj in https://github.com/Blaizzy/mlx-vlm/pull/1369
[Fix] interleaved reasoning by @Terrencezzj in https://github.com/Blaizzy/mlx-vlm/pull/1370
Fix color shift from tiled VAE in Flux.2 models by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1381
Revert "[Cohere] Fix streaming thinking parsing after prompt-opened" by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1389
Fix Gemma 4 namespaced tool calls by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1379
Forward strict flag when loading weights by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1345
Add optional API key for the server by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1385
Fix SFT masks for Gemma4 unified by @kqw8 in https://github.com/Blaizzy/mlx-vlm/pull/1392
Add preloading for multiple model kinds in the server by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1372
fixes #1404: wire --prefill-step-size into BatchGenerator constructor by @royteeuwen in https://github.com/Blaizzy/mlx-vlm/pull/1405
fixes #1402: Every first request after server startup causes a full model reload by @royteeuwen in https://github.com/Blaizzy/mlx-vlm/pull/1403
fixes #1406: move draft-kind/draft-block-size env vars outside draft-model guard by @royteeuwen in https://github.com/Blaizzy/mlx-vlm/pull/1407
fix(StoppingCriteria): copy eos_token_ids to prevent caller-list aliasing by @raullenchai in https://github.com/Blaizzy/mlx-vlm/pull/1410
[performance] Optimize PaddleOCR-VL vision fast paths by @scyyh11 in https://github.com/Blaizzy/mlx-vlm/pull/1396
Avoid clearing MLX cache from stream finalizers by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1412
Add Moondream2 model support by @Abdullah-Alharbi61 in https://github.com/Blaizzy/mlx-vlm/pull/1395
Qwen3.5 on CUDA: fix Metal-only crashes + chunked gated-delta prefill (8.5x faster prefill) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1423
Add reasoning_content to chat messages by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1411
improved turbo quant's prefill by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1433
adding hisa to dsa based models(currently only deepseek) by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1434
fix for #1446 by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1447
Fix Qwen3.5 MoE gate-up sanitization by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1450
Fix Gemma 4 cached vision mask overlay by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1445
Fix stale GenerationBatch processor state by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1451
kimi_k25: 3D MoonViT video support (Fixes #1436) by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1438
Fix TurboQuant MTP rollback tail zeroing by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1432
convert: operate on the inner model for text-only wrappers (Fixes #1437) by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1439
Fix Qwen3.5 TurboQuant batch rollback trim by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1453
Fix quantized Diffusion Gemma sampler speed by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1359
Remove video generate command by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1454
fix(StoppingCriteria): copy eos_token_ids in reset() to prevent caller-list aliasing by @SuperMarioYL in https://github.com/Blaizzy/mlx-vlm/pull/1430
Guard empty MRoPE position batches by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1455
Improve Responses API reasoning and tool stream events by @justin-schroeder in https://github.com/Blaizzy/mlx-vlm/pull/1458
Select weight files via model.safetensors.index.json when present(fix #1329) by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1464
fix(apc): enable APC prefix cache with MTP speculative decoding by @vedang in https://github.com/Blaizzy/mlx-vlm/pull/1463
fix(deepseek_v4): clamp shared-expert SwiGLU to swiglu_limit by @popfido in https://github.com/Blaizzy/mlx-vlm/pull/1469
Fix Gemma 4 bare namespaced tool calls by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1473
fix(qwen3_5_moe): support split expert weights for ornith 1.0 35b by @scaryrawr in https://github.com/Blaizzy/mlx-vlm/pull/1472
Minimax m3 support by @ivanfioravanti in https://github.com/Blaizzy/mlx-vlm/pull/1374
Add Unlimited-OCR model support by @shuuul in https://github.com/Blaizzy/mlx-vlm/pull/1427
fix(qwen3-omni): eval mRoPE position_ids before the thinker forward by @yuhuanowo in https://github.com/Blaizzy/mlx-vlm/pull/1461
fix(rope_utils): make M-RoPE apply differentiable for training by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1475
Fix Qwen3-Omni repeated generation by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1476
Fix DiffusionGemma tool marker tokenization by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1477
Constrain transformers below 5.13 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1488
fix #1480 by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1484
Make Qwen MRoPE request-local under continuous batching by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1486
fix #1481: quantized KV cache crashes in mask-length checks by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1483
Reject empty OpenAI server inputs by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1491
Consolidate caches + RoPE handling by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1494
fix(lfm2_vl): only create projector LayerNorm when projector_use_laye… by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1493
Bump version to 0.6.4 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1499
Fix MolmoPoint EOS config by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1500
Fix Phi-3-V chat stop tokens by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1501
Always sanitize loaded weights and simplify audio preprocessing by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1498
Add torch-free Pixtral processor by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1502
Support Gemma 4 variable soft-token budgets by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1503
Fix DiffusionGemma processor and video inputs end-to-end by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1492
Fix MiniCPM-V 4.6 weight sanitization by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1506
Use model type in diffusion text-only error by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1507
Unify diffusion generation path by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1508
migration from mlx-lm dependency by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1509
Remove obsolete test mocks for safetensors by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1511
Port glm_moe_dsa language model by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1390
Fix computer use action parsing by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1514
Add qwen2 model type and remove mlx-lm from fastvlm by @Lazarus-931 in https://github.com/Blaizzy/mlx-vlm/pull/1516

New Contributors

@imi4u36d made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1354
@kqw8 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1392
@royteeuwen made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1405
@raullenchai made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1410
@Abdullah-Alharbi61 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1395
@Lazarus-931 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1433
@justin-schroeder made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1458
@vedang made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1463
@scaryrawr made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1472
@shuuul made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1427

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.6.3...v0.6.4

Open source →

0.6.310 Jun 2026

Fix APC splitting on multimodal inputs by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1311

What's Changed

Fix APC splitting on multimodal inputs by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1311
Fix Qwen quantized KV cache prompt state by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1313
Fix Gemma 4 unified silently dropping image/video inputs (video processor rejects standard HF config keys) by @sockeye44 in https://github.com/Blaizzy/mlx-vlm/pull/1321
Fix thinking enabled by default for chat templates by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1316
Fix APC exact disk-hit entries not promoted to in-memory LRU by @mikeatlas in https://github.com/Blaizzy/mlx-vlm/pull/1324
Align Qwen3-VL visual masks to the current prefill / decode window by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1325
Fix LFM2.5 VL model loading by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1328
Fix Phi 3.5 VL eos tokens by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1326
Fix Qwen3-VL deepstack visual embeds misaligned during chunked prefill by @sockeye44 in https://github.com/Blaizzy/mlx-vlm/pull/1332
Fix --system prompt ignored by one-shot generate by @sockeye44 in https://github.com/Blaizzy/mlx-vlm/pull/1330
Improve speculative prefill when using MTP drafters by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1334
Respect generation_config sampling defaults by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1337
Add server thinking defaults by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1342
Add Gemma 4 DLM by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1347
Fix DiffusionGemma long-context prefill by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1348

New Contributors

@sockeye44 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1321
@mikeatlas made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1324

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.6.2...v0.6.3

Open source →

0.6.25 Jun 2026

Make sure to evaluate cached MRoPE parameters on model load by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1272

What's Changed

Make sure to evaluate cached MRoPE parameters on model load by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1272
Fix Gemma4 unified long text prefill by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1280
Add support for the Ideogram 4 image generation model by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1281
Fix Nemotron-H Nano Omni audio crash under nvfp4 quantization by @txdadlab in https://github.com/Blaizzy/mlx-vlm/pull/1279
Fix multi-image prompts in get_rope_index for remaining MRoPE families by @scyyh11 in https://github.com/Blaizzy/mlx-vlm/pull/1282
Fix PaddleOCR-VL stale mRoPE state under batching by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1285
Add Ideogram 4 local prompt expansion support by @omercelik in https://github.com/Blaizzy/mlx-vlm/pull/1276
Fix APC for single requests by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1291
Add video input support for Gemma 4 12B by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1292
Fix stale test_utils.py regressions + extract get_class_predicate by @mdstaff in https://github.com/Blaizzy/mlx-vlm/pull/1071
Add regression tests for ThinkingBudgetCriteria.apply_forced_token by @glenncameron2 in https://github.com/Blaizzy/mlx-vlm/pull/1297
[Cohere] Add Cohere tool-calling and thinking parsing support by @Terrencezzj in https://github.com/Blaizzy/mlx-vlm/pull/1298
docs(gemma4): clarify base vs instruct and thinking modes by @Girish011 in https://github.com/Blaizzy/mlx-vlm/pull/1300
test: add PaddleOCR-VL processor regression coverage by @jimmyzhuu in https://github.com/Blaizzy/mlx-vlm/pull/933
docs: add model guides for ERNIE 4.5 VL and PaddleOCR-VL by @jimmyzhuu in https://github.com/Blaizzy/mlx-vlm/pull/934
Fix using structured output with thinking enabled by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1299
Accept qwen3_5_vision / qwen3_5_moe_vision vision model_type by @davidrhodus in https://github.com/Blaizzy/mlx-vlm/pull/1302
Don't specify a default model for the server by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1303
Fix gemma4 load for HF checkpoints with num_kv_shared_layers > 0 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1301
Update version to v0.6.2 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1304
Adjust Gemma4 quantization predicate by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1288
Fix LFM2-VL checkpoint loading by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1308
Fix Kimi VL quantized projector loading by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1309

New Contributors

@txdadlab made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1279
@scyyh11 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1282
@glenncameron2 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1297
@Girish011 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1300
@jimmyzhuu made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/933
@davidrhodus made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1302

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.6.1...v0.6.2

Open source →

0.6.13 Jun 2026

Move linear spec handling to model backends by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1259

What's Changed

Move linear spec handling to model backends by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1259
Add NVIDIA LocateAnything-3B (MoonViT + Qwen2.5, AR + Parallel Box Decoding) by @beshkenadze in https://github.com/Blaizzy/mlx-vlm/pull/1242
Fix Gemma 4 MTP rollback crash when accepted is a list (#1260) by @francip in https://github.com/Blaizzy/mlx-vlm/pull/1261
[Cohere] Add cohere2_moe model support by @Terrencezzj in https://github.com/Blaizzy/mlx-vlm/pull/1268
Fix system role normalization for /v1/messages API by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1269
Add Gemma 4 Unified + MTP support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1267
Fix Gemma 4 rollback handling and streaming thinking splits by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1266

New Contributors

@francip made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1261
@Terrencezzj made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1268

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.6.0...v0.6.1

Open source →

0.6.01 Jun 2026Deprecated

fix: replace deprecated use_fast=True with backend=torchvision in load_processor by @Jonathangadeaharder in https://github.com/Blaizzy/mlx-vlm/pull/11…

What's Changed

fix(server): remove duplicate “server” response header by @goniz in https://github.com/Blaizzy/mlx-vlm/pull/1140
Fix top_p_sampling crash on 3D logits in speculative decoding by @ivanfioravanti in https://github.com/Blaizzy/mlx-vlm/pull/1141
Use local rotating-cache offsets for Gemma assistant MTP masks by @Thump604 in https://github.com/Blaizzy/mlx-vlm/pull/1139
Fix stale Qwen MRoPE rope_deltas in decode by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/1143
Fix server chat completions text content array handling by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/1137
Add server metrics and improve streamed response handling by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1145
Fix streamed byte-fallback token handling in server batching by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1147
Fix APC disk cache directory recovery by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1149
minicpmo / fastvlm: fix pixel cast on quantized language models by @contrapuntal in https://github.com/Blaizzy/mlx-vlm/pull/1098
Fix Qwen3 VL vision config deserialization (patch_size ignored) by @lyonsno in https://github.com/Blaizzy/mlx-vlm/pull/1157
Fix max kv size enforcement by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1160
[codex] Enforce server context budget by @eloe in https://github.com/Blaizzy/mlx-vlm/pull/1153
Fix LFM2-VL model by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1162
Add ZAYA1-VL model support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1159
Fix an error when loading the MolmoPoint-8B model by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1164
Fix image feature fp16 overflow in Molmo2 by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1165
Improve Gemma4 MTP server batching by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1166
Add Qwen MTP speculative decoding by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1167
Add MiniCPM-V 4.6 support by @pzc163 in https://github.com/Blaizzy/mlx-vlm/pull/1058
Refactor speculative decoding utilities by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1169
Fix crash with FastAPI workers when using Qwen3.5 family models by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1168
Add Gemma 4 DFlash + adaptive block sizing by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1176
fix: replace deprecated use_fast=True with backend=torchvision in load_processor by @Jonathangadeaharder in https://github.com/Blaizzy/mlx-vlm/pull/1170
Fix weight remapping for the Qwen3.5/3.6 vision tower by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1179
Fix Nemotron processor metadata loading by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1177
Add Gemma 4 EAGLE3 speculative decoding support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1180
Split speculative decoding utilities by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1182
Add Laguna support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1183
Compatibility bridge for non-VL models by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1181
Unify LoRA loading for VLM/LM models by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1186
[mlx_vlm] Expose 'strict' parameter in load() function by @zyguy in https://github.com/Blaizzy/mlx-vlm/pull/1198
Add server support for Anthropic-style /v1/messages API by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1196
Support /v1/responses stateful responses API in the server by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1199
Fix /v1/messages tool call image handling by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1200
Fix Qwen MTP verification drift and sampling parity by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1188
Modularize server by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1203
Fix mask creation for batched quantized kv caches by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1208
Track APC cached tokens for generation metrics by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/1209
fastvlm: support untied lm_head for the 7B variant by @contrapuntal in https://github.com/Blaizzy/mlx-vlm/pull/1193
Add RT-DETRv2 detection model by @leonnoirclerc in https://github.com/Blaizzy/mlx-vlm/pull/1195
Handle base64-encoded input audio in chat completions by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1211
Fix inverted --verbose flag on chat / generate / video_generate CLIs by @SuperMarioYL in https://github.com/Blaizzy/mlx-vlm/pull/1152
Fix for Gemma4 audio task by @lifeiteng in https://github.com/Blaizzy/mlx-vlm/pull/980
Add MiniCPM-V 4.6 language wrapper by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1212
Support Qwen image pixel overrides by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1213
Improve GLM-OCR generation and repetition controls by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1214
Fix LFM2-VL image preprocessing for variants shipping input_data_format=channels_last by @contrapuntal in https://github.com/Blaizzy/mlx-vlm/pull/1190
Fix rotation-basis mismatch in TurboQuant L=1 fast quantize by @nnorris7 in https://github.com/Blaizzy/mlx-vlm/pull/973
fix: fallback to defaults by @GeneCodeSavvy in https://github.com/Blaizzy/mlx-vlm/pull/1221
Refactor MRoPE handling by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1135
Fix server finish_reason inconsistencies by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/1215
Fix inference cancellation on stream abort by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/1217
Add DeepSeek V4 language model by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1223
Add spec-compliant usage and timings for server endpoints by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/1216
Add DeepSeek V4 MTP support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1225
Freeze MaskedEmbedder.token_ordering so optimizer can't corrupt it by @DirectriX01 in https://github.com/Blaizzy/mlx-vlm/pull/1194
Validate MTP drafter compatibility by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1227
Restructure generation into AR and diffusion engines by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1229
Add support for image generation with PrismML Bonsai by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1226
Enforce thinking budget in server batching by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1228
Add LLaDA2.X support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1230
Add support for FLUX.2 base & klein models by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1232
Improve image generation quality for Bonsai by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1234
Add support for image editing with FLUX.2 models by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1236
Improve masked diffusion text generation controls by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1235
Add HRM-Text model support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1238
Add LFM2 MoE language model by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1237
Add Step-3.7 Flash support by @ivanfioravanti in https://github.com/Blaizzy/mlx-vlm/pull/1245
fix(kernels): match PyTorch bicubic coefficient (-0.75 non-AA / -0.5 AA) — closes #1241 by @beshkenadze in https://github.com/Blaizzy/mlx-vlm/pull/1243
Fix(qwen3_omni_moe): enable audio input (mask key + audio placeholder) by @yuhuanowo in https://github.com/Blaizzy/mlx-vlm/pull/1240
docs: align Python version metadata by @aqilaziz in https://github.com/Blaizzy/mlx-vlm/pull/1150
Fix ByteLevel streaming detokenizer detection for Step 3.7 Flash by @ivanfioravanti in https://github.com/Blaizzy/mlx-vlm/pull/1246
turboquant: guard L=1 value kernels behind not use_rht (fix masked decode under RHT) by @popfido in https://github.com/Blaizzy/mlx-vlm/pull/1244
Fix cache merging for batched quantized caches by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1248
security: require patched Starlette by @Thump604 in https://github.com/Blaizzy/mlx-vlm/pull/1249
Don't show tqdm progress bars for prefill by default by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1250
Fix adaptors with the /v1/responses API by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1251
[codex] Fix models endpoint for loaded local models by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1253
perf(turboquant): RHT-correct L=1 value kernels (keep the fast path under RHT) by @popfido in https://github.com/Blaizzy/mlx-vlm/pull/1252
Fix trainer adapter save fallback by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1255
Fix Qwen MTP batched target-verify drift by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1210
Add Nemotron Labs Diffusion model by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1239
Update version to v0.6.0 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1257

New Contributors

@contrapuntal made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1098
@lyonsno made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1157
@pzc163 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1058
@Jonathangadeaharder made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1170
@zyguy made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1198
@leonnoirclerc made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1195
@SuperMarioYL made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1152
@lifeiteng made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/980
@nnorris7 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/973
@GeneCodeSavvy made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1221
@DirectriX01 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1194
@beshkenadze made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1243
@yuhuanowo made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1240
@aqilaziz made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1150
@popfido made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1244

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.5.0...v0.6.0

Open source →

0.5.06 May 2026

Fix gemma4 multi-image processing for different-sized images by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/938

What's Changed

Fix gemma4 multi-image processing for different-sized images by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/938
Fix race condition in TurboQuant fused fast-quantize kernels by @Chedrian07 in https://github.com/Blaizzy/mlx-vlm/pull/967
Fix Gemma 4 tool parser to accept hyphenated function names by @michaelstingl in https://github.com/Blaizzy/mlx-vlm/pull/963
fix(gemma4): snapshot cache.offset to prevent alias mutation under batched caches by @Thump604 in https://github.com/Blaizzy/mlx-vlm/pull/966
Fix Gemma 4 quantized per-layer projection loading by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/935
fix: Gemma 4 audio — mel preprocessing, weight loading, feature extractor by @stephencox-ict in https://github.com/Blaizzy/mlx-vlm/pull/931
Add continuous batching to server by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1027
fix: use alpha/rank scaling in LoRaLayer (standard LoRA convention) by @kikoncuo in https://github.com/Blaizzy/mlx-vlm/pull/846
Fix duplicate docstring entries and minor issues in utils.py by @sjhddh in https://github.com/Blaizzy/mlx-vlm/pull/925
Add KV cache quantization for continuous batching by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1030
Strip tool-call markup from streamed delta.content by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1037
Add DFlash speculative decoding (single + batch + server) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1029
Fix stale position IDs and gdn_sink compatibility for Qwen3.5/3.6 MoE… by @lele872 in https://github.com/Blaizzy/mlx-vlm/pull/1040
Distributed inference for Qwen3, Kimi K2.5 and K2.6 by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/689
Add Distributed Inference section by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/1041
Resolve no images crash for qwen3_vl and qwen3_vl_moe generate call by @urimem in https://github.com/Blaizzy/mlx-vlm/pull/1013
Add Youtu-VL by @MollySophia in https://github.com/Blaizzy/mlx-vlm/pull/1018
Fix preprocessing for image input for trainer by @Goekdeniz-Guelmez in https://github.com/Blaizzy/mlx-vlm/pull/826
add grounded_reasoning: Falcon Perception + Gemma4 agentic demo by @YasserdahouML in https://github.com/Blaizzy/mlx-vlm/pull/926
Add Gemma 4 video support (multi-video capable) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1042
Qwen2 / 2.5 / 3 / 3.5 VL: torch-free video processors + chunked-prefill rope fix by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1048
Close the batch_generate / server decode gap + VLM fixes by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1055
Thread-local generation stream (port mlx-lm#1090) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1050
Fix DFlash speculative decoding: GPU hang, performance, and upstream alignment by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1053
hunyuan_vl / gemma3n: drop dead assignments in cache-offset extraction by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1056
Fix Gemma 4 LoRA training: vision backward NaN + audio_tower freeze leak by @john-rocky in https://github.com/Blaizzy/mlx-vlm/pull/1052
server: added 'server' header in responses by @goniz in https://github.com/Blaizzy/mlx-vlm/pull/1082
Add Nemotron 3 Nano Omni model by @lucasnewman in https://github.com/Blaizzy/mlx-vlm/pull/1087
Nemotron H Nano Omni: rename SoundConfig and add processor by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1088
mistral3: skip lm_head quantization and add multi_modal_projector to skip list by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1089
Add server json_schema response_format support by @avbiswas in https://github.com/Blaizzy/mlx-vlm/pull/1047
Fix Kimi VL concurrent Metal crash and mixed-batch text degradation by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1039
Fix per-sequence MRoPE alignment in mixed VL+text batches by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/1095
granite4_vision: support standard granite MLP backbone for 4.1-4b by @EliSchwartz in https://github.com/Blaizzy/mlx-vlm/pull/1104
Add Gemma 4 MTP speculative-decoding drafter by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1112
Add APC prompt caching with disk persistence by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1103
Add APC prompt caching with warm-disk persistence for hybrid models by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1114
Fix token queue timeout during long prefills by @eloe in https://github.com/Blaizzy/mlx-vlm/pull/1111
Server: add Gemma 4 MTP drafter support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1115
Fix TurboQuant batch cache offset merging by @eloe in https://github.com/Blaizzy/mlx-vlm/pull/1110
Fix Qwen3.5 quantization config keys by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1119
server: added loaded model's context size and tool call parser to /health endpoint by @goniz in https://github.com/Blaizzy/mlx-vlm/pull/1092
fix: Batch generation breaks top-p sampling by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/1094
feat: Add SAM 3D Body — monocular 3D body mesh on Apple Silicon by @shihwesley in https://github.com/Blaizzy/mlx-vlm/pull/922
Add --max-tokens to server by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/1120
refactor: improve model loading and resource handling in utils.py by @SyedaAnshrahGillani in https://github.com/Blaizzy/mlx-vlm/pull/1019
Fix Gemma 4 E2B/E4B per_layer_inputs crash in batched prefill by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1123
Fix mixed-length Gemma 4 batching by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1127
Fix streamed detokenization for byte-fallback tokens by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1129
Add server thinking mode flag by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1130
Speculative decoding fixes: auto-detect drafter kind and preserve multimodal prefill by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/1125

New Contributors

@Chedrian07 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/967
@michaelstingl made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/963
@Thump604 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/966
@stephencox-ict made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/931
@kikoncuo made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/846
@sjhddh made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/925
@lele872 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1040
@urimem made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1013
@MollySophia made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1018
@YasserdahouML made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/926
@john-rocky made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1052
@goniz made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1082
@lucasnewman made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1087
@avbiswas made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1047
@EliSchwartz made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1104
@eloe made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1111
@shihwesley made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/922
@SyedaAnshrahGillani made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/1019

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.4.4...v0.5.0

Open source →

0.4.44 Apr 2026

Fix Gemma 4 chunked prefill for KV-shared models and thinking by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/901

What's Changed

Fix Gemma 4 chunked prefill for KV-shared models and thinking by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/901
Fix Gemma 4 vision + text degradation and missing processor config by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/906
Fix Falcon-Perception 300M and move generate_perception to model by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/910
Fix Gemma 4 tool parser for nested arguments by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/916
Add VisionFeatureCache for multi-turn image caching by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/913
Fix broken video_generate and smolvlm_video_generate CLI commands by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/919
Optimize TurboQuant Metal kernels: 0.85-1.90x baseline with 89% KV savings by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/909

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.4.3...v0.4.4

Open source →

0.4.32 Apr 2026

Add SAM 3.1 with Object Multiplex and optimized realtime pipeline by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/880

What's Changed

Add SAM 3.1 with Object Multiplex and optimized realtime pipeline by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/880
Add Falcon-OCR model support by @Griffintaur in https://github.com/Blaizzy/mlx-vlm/pull/879
Add RF-DETR detection and segmentation model by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/884
Add granite vision 3.2 and 4.0 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/885
Add wired_limit to vision model inference pipelines for CUDA support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/887
Add falcon perception by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/888
fix(server): disable uvicorn reload by default to prevent memory leaks by @futurepitcher in https://github.com/Blaizzy/mlx-vlm/pull/883
Add Gemma 4 model support (vision, audio, MoE) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/890
Fix Gemma 4 embedding scaling by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/893
Add Turbo Quant by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/858
Remove TurboQuant benchmark artifacts and add README docs by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/894

New Contributors

@Griffintaur made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/879
@futurepitcher made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/883

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.4.2...v0.4.3

Open source →

0.4.228 Mar 2026

Ensure minimal metadata exists in model card by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/853

What's Changed

Ensure minimal metadata exists in model card by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/853
fix(qwen3_5): register AutoProcessor patch for qwen3_5 model type by @mdstaff in https://github.com/Blaizzy/mlx-vlm/pull/859
Patch qwen3_5_moe auto processor by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/867
Cast Qwen3.5 RMSNorm output to input dtype by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/868
Fix thinking defaults in CLI and server by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/869
Fix LFM2-VL processor loading without torch by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/872
Fix Magistral image token expansion by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/873
Add Facebook Sam3 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/875
[Sam3] label drawing for mask only in realtime video by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/876
Add DOTS-MOCR processor support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/874
Fix PaliGemma processor kwarg routing by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/877

New Contributors

@mdstaff made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/859

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.4.1...v0.4.2

Open source →

0.4.121 Mar 2026

feat(server): add --model and --adapter-path flags for startup preloading by @auggie246 in https://github.com/Blaizzy/mlx-vlm/pull/811

What's Changed

feat(server): add --model and --adapter-path flags for startup preloading by @auggie246 in https://github.com/Blaizzy/mlx-vlm/pull/811
fix(qwen3_5): guard batch dimension mismatches under continuous batching by @kol22 in https://github.com/Blaizzy/mlx-vlm/pull/813
Add common sampling parameters to server endpoints by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/814
Add an 'index' for each tool call as per OpenAI spec. by @viktike in https://github.com/Blaizzy/mlx-vlm/pull/818
Fix Qwen3-Omni (qwen3_omni_moe) integration by @howeirdo in https://github.com/Blaizzy/mlx-vlm/pull/820
Add mistral4 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/827
Clean up server + generate args, validation, and defaults by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/829
Set custom processor as default and remove torch as dep by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/821
Add molmo point by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/844

New Contributors

@auggie246 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/811
@howeirdo made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/820

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.4.0...v0.4.1

Open source →

0.4.07 Mar 2026

Fix gemma3n short prompts by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/751

What's Changed

Fix gemma3n short prompts by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/751
Adding full weight finetuning by @Goekdeniz-Guelmez in https://github.com/Blaizzy/mlx-vlm/pull/499
Initialize qwen3_5_moe.LanguageModel with _position_ids by @will-lms in https://github.com/Blaizzy/mlx-vlm/pull/753
Fix batch inference to use InputEmbeddingsFeatures by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/760
fix(qwen3_vl): reset _position_ids when new image/video arrives by @kol22 in https://github.com/Blaizzy/mlx-vlm/pull/756
fix: replace 13 bare excepts with except Exception by @haosenwang1018 in https://github.com/Blaizzy/mlx-vlm/pull/765
Reset position ids for Qwen3-VL-MoE image inputs by @kol22 in https://github.com/Blaizzy/mlx-vlm/pull/761
Fix text only single batch by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/771
Update LORA.md for the new trainer by @Goekdeniz-Guelmez in https://github.com/Blaizzy/mlx-vlm/pull/775
Support /v1/ prefix for OpenAI compatible endpoints by @spicyneuron in https://github.com/Blaizzy/mlx-vlm/pull/783
CORS middleware by @viktike in https://github.com/Blaizzy/mlx-vlm/pull/766
fx: README curl error by @yinzhidong in https://github.com/Blaizzy/mlx-vlm/pull/788
Streaming chatml response enhancements. by @viktike in https://github.com/Blaizzy/mlx-vlm/pull/764
Add KV cache (quantization) parameters to server. by @viktike in https://github.com/Blaizzy/mlx-vlm/pull/776
Fix mllama training and add more models in mapp by @Goekdeniz-Guelmez in https://github.com/Blaizzy/mlx-vlm/pull/777
Add thinking budget and flag by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/789
fix: add import guard for gradio in chat_ui entry point by @RtYkk in https://github.com/Blaizzy/mlx-vlm/pull/787
Tool calling in server by @viktike in https://github.com/Blaizzy/mlx-vlm/pull/773
Add Minicpm-o-2.5 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/791
Add prefill-step-size optional command line argument for server. by @viktike in https://github.com/Blaizzy/mlx-vlm/pull/792
qwen3_omni_moe: fix visual_embeds_multiscale UnboundLocalError by @ronaldseoh in https://github.com/Blaizzy/mlx-vlm/pull/794
Add Phi-4-reasoning-vision-15B (phi4-siglip) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/796
Add phi4mm by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/797
Add ensure_fused_sdpa function for optimized attention computation by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/800
[Internvl_chat] Add optional kwargs parameter to LanguageModel by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/801
Guard load img & audio by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/803
[Qwen3.5 MoE] Fix quant predicate by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/804
Adding orpo by @Goekdeniz-Guelmez in https://github.com/Blaizzy/mlx-vlm/pull/795
Add moondream3 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/807
Remove sleep duration with zero delay by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/808
Update dependencies and version number by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/809

New Contributors

@kol22 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/756
@haosenwang1018 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/765
@spicyneuron made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/783
@viktike made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/766
@yinzhidong made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/788
@RtYkk made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/787
@ronaldseoh made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/794

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.12...v0.4.0

Open source →

0.3.1216 Feb 2026

[MODEL] support qwen3.5 series by @JJJYmmm in https://github.com/Blaizzy/mlx-vlm/pull/722

What's Changed

[MODEL] support qwen3.5 series by @JJJYmmm in https://github.com/Blaizzy/mlx-vlm/pull/722
qwen3_omni_moe: Fix destructuring error by @Nixuge in https://github.com/Blaizzy/mlx-vlm/pull/723
docs: clarify that some models need [torch] extra for torchvision by @Mr-Neutr0n in https://github.com/Blaizzy/mlx-vlm/pull/716
Use mlx-lm provided logits processors and samplers by @AG6GR in https://github.com/Blaizzy/mlx-vlm/pull/724
Fix LoRA training crash when config uses image_token_id by @JesseRod329 in https://github.com/Blaizzy/mlx-vlm/pull/720
video_generate.py: handle capitalized extensions in is_video_file by @Nixuge in https://github.com/Blaizzy/mlx-vlm/pull/719
Honor quantization_config by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/692
[Ministral] Add FP8 dequantization for model weights by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/727
Fix LFM2.5VL processor compatibility patch image placeholder count by @mattjcly in https://github.com/Blaizzy/mlx-vlm/pull/725
Add activation quantization support for QQLinear layers by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/728
Refactor processor return types in utils.py to use ProcessorMixin by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/729
Fix glm (4v & 4v moe) broadcast by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/731
[FastVLM] fix dtype cast by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/733
Fix kimi broadcast by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/734
Add custom processor for fastvlm by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/736
Fix paligemma prefill and numerical precision by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/737
Fix idefics3 llama3 and smovlm by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/738
docs(README): fix JSON formatting in curl examples by @chsdwn in https://github.com/Blaizzy/mlx-vlm/pull/739
fix chunk prefill in qwen3.5 moe by @JJJYmmm in https://github.com/Blaizzy/mlx-vlm/pull/742
Fix florence-2 (processor and infer) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/743
Fix Qwen3.5 cast type and predicates by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/744
[Qwen2, 2.5] Fix vision overflow by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/745
[Ministral3] Fix multi-image by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/747
[Phiv3] Fix dtype cast by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/748
Add dots-ocr by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/749

New Contributors

@Nixuge made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/723
@Mr-Neutr0n made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/716
@AG6GR made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/724
@JesseRod329 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/720
@chsdwn made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/739

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.11...v0.3.12

Open source →

0.3.114 Feb 2026

Refactor input embedding handling in _generate_batch function by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/694

What's Changed

Refactor input embedding handling in _generate_batch function by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/694
[Qwen2-VL ] Fix incorrect attention mask in Vision by @hturbe in https://github.com/Blaizzy/mlx-vlm/pull/704
Add GLM-OCR by @Blaizzy & @mikolaj92 in https://github.com/Blaizzy/mlx-vlm/pull/706
fix: Log underlying ImportError when model loading fails by @antonvice in https://github.com/Blaizzy/mlx-vlm/pull/710
Fix wired limit, deepstack eval and add prefill step size argument by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/699
[PaddleOCR] Fix hardcoded processor config by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/712
[SmolVLM] Refactor Attention class to calculate n_kv_heads dynamically by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/713

New Contributors

@hturbe made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/704
@mikolaj92 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/706
@antonvice made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/710

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.10...v0.3.11

Open source →

0.3.1028 Jan 2026

Fix qk_norm for lighton by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/615

What's Changed

Fix qk_norm for lighton by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/615
Fix GLM-4.6V vision types by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/616
Add support for rope parameters [GLM-4.6V MoE] by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/617
Enhance chat UI by @ivanfioravanti in https://github.com/Blaizzy/mlx-vlm/pull/619
Add support for TokenizersBackend by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/621
UI by @ivanfioravanti in https://github.com/Blaizzy/mlx-vlm/pull/620
docs: fix /chat/completion endpoint typo in README by @zenyr in https://github.com/Blaizzy/mlx-vlm/pull/623
Add Support for HunyuanOCR (hunyuan_vl) by @voxmenthe in https://github.com/Blaizzy/mlx-vlm/pull/604
Add Batch Generation by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/538
Add autoflake (remove unused imports) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/628
Bump version to 0.3.10 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/632
Fix qwen3_vl_moe kwargs forwarding by @mattjcly in https://github.com/Blaizzy/mlx-vlm/pull/633
Add support for Jina VLM by @hanxiao in https://github.com/Blaizzy/mlx-vlm/pull/631
Fix: Skip index.json in file copy to prevent overwrite by @mzau in https://github.com/Blaizzy/mlx-vlm/pull/638
Add Molmo2 support by @aliyovic in https://github.com/Blaizzy/mlx-vlm/pull/639
Fix glm46v by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/629
migrate qwen3omni to MLX by @hellopahe in https://github.com/Blaizzy/mlx-vlm/pull/598
Fix audio loading by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/642
Fix pixtral text-only regression from batch generation PR by @mzau in https://github.com/Blaizzy/mlx-vlm/pull/644
Fix: fixes mlx-vlm.generate --chat when prompted with no imgages by @Deekshith-Dade in https://github.com/Blaizzy/mlx-vlm/pull/648
Enhance load_config to include generation_config.json and extract eos… by @cubist38 in https://github.com/Blaizzy/mlx-vlm/pull/650
Add LFM2.5-VL by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/653
Add MXFP4 Quantization Support by @zhutao100 in https://github.com/Blaizzy/mlx-vlm/pull/514
Add support for nvfp4 and mxfp8 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/657
Fix server audio, chat and trust remote code by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/660
Fix tokenizer by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/662
Pass kwargs through to snapshot_download in get_model_path by @Deekshith-Dade in https://github.com/Blaizzy/mlx-vlm/pull/663
Revert "Pass kwargs through to snapshot_download in get_model_path" by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/666
[Model] Add PaddleOCR-VL Model Support by @zhang-prog in https://github.com/Blaizzy/mlx-vlm/pull/656
Remove _has_bytelevel_pretokenizer by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/668
Feat/structured outputs by @cubist38 in https://github.com/Blaizzy/mlx-vlm/pull/664
Fix: Correct position embedding for LLM in PaddleOCR-VL by @zhang-prog in https://github.com/Blaizzy/mlx-vlm/pull/672
Use mx.fast.scaled_dot_product_attention in pixtral vision by @pherber3 in https://github.com/Blaizzy/mlx-vlm/pull/677
fix: Remove invalid parameter "None", explicitly specify assignment p… by @neil0306 in https://github.com/Blaizzy/mlx-vlm/pull/687
Add text prefill and input embeddings obj by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/681
Make sure weights are released during conversion by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/691
Add deepseek ocr 2 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/690
fix: prevent base64 image data from being tokenized as text by @ne2030 in https://github.com/Blaizzy/mlx-vlm/pull/685
Add Ernie-4.5-VL by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/627
TFMS v5 RC3 + Fix processor registry by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/693

New Contributors

@voxmenthe made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/604
@hanxiao made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/631
@mzau made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/638
@aliyovic made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/639
@hellopahe made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/598
@Deekshith-Dade made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/648
@cubist38 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/650
@zhutao100 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/514
@zhang-prog made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/656
@pherber3 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/677
@neil0306 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/687
@ne2030 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/685

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.9...v0.3.10

Open source →

0.3.93 Dec 2025

Fix qwen3_vl ValueError: Image features and image tokens do not match by @ziya32 in https://github.com/Blaizzy/mlx-vlm/pull/608

What's Changed

Fix qwen3_vl ValueError: Image features and image tokens do not match by @ziya32 in https://github.com/Blaizzy/mlx-vlm/pull/608
Add ministral3 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/611

New Contributors

@ziya32 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/608

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.8...v0.3.9

Open source →

0.3.827 Nov 2025

Add chat_ui command by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/589

What's Changed

Add chat_ui command by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/589
Fix 422 error when calling /chat/completions with OpenAI SDK #590 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/592
Fix qwen3_vl by @awni in https://github.com/Blaizzy/mlx-vlm/pull/599

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.7...v0.3.8

Open source →

0.3.717 Nov 2025

update readme with new openai endpoints details by @mguella in https://github.com/Blaizzy/mlx-vlm/pull/585

What's Changed

update readme with new openai endpoints details by @mguella in https://github.com/Blaizzy/mlx-vlm/pull/585
Fix lighton OCR empty result on main by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/587
Add model revision loading by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/588

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.6...v0.3.7

Open source →

0.3.614 Nov 2025

Fix: Input cast error by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/559

What's Changed

Fix: Input cast error by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/559
Fix qwen-vl position ids (2.5 & 3) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/564
Add Evals by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/563
Fix Qwen3-VL Attention by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/566
Update evals init by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/567
Fix Qwen3-VL multi image reshape by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/569
Add processor args to DeepSeekOCR by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/570
host and port params for server by @mguella in https://github.com/Blaizzy/mlx-vlm/pull/568
FastVLM by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/502
Add support for InternVL3 by @iRonJ in https://github.com/Blaizzy/mlx-vlm/pull/540
Add Z.ai GLM-4.1v by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/572
Make rope deltas private by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/573
Add example notebook for interleaving text and images in prompts by @Copilot in https://github.com/Blaizzy/mlx-vlm/pull/574
[Bugfix] fix mrope in qwen2vl and qwen2.5vl by @JJJYmmm in https://github.com/Blaizzy/mlx-vlm/pull/576
Add lighton-ocr by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/550
changed image parameter instead of files in stream_generate by @Manikandan-t in https://github.com/Blaizzy/mlx-vlm/pull/521
Remove auto config loading by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/577
[BugFix][Qwen3VL] fix deepstack and multi-image inference by @JJJYmmm in https://github.com/Blaizzy/mlx-vlm/pull/581
openai compatible endpoints by @mguella in https://github.com/Blaizzy/mlx-vlm/pull/580
Bump version to 0.3.6 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/582

New Contributors

@mguella made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/568
@iRonJ made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/540
@Copilot made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/574
@JJJYmmm made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/576
@Manikandan-t made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/521

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.5...v0.3.6

Open source →

0.3.526 Oct 2025

Remove docs actions temporarly by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/536

What's Changed

Remove docs actions temporarly by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/536
Fix load_image for URLs by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/534
Add Deepseek ocr by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/541

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.4...v0.3.5

Open source →

0.3.414 Oct 2025

Add n_kv_heads property used in LM Studio to glm4v_moe.LanguageModel by @hehua2008 in https://github.com/Blaizzy/mlx-vlm/pull/472

What's Changed

Add n_kv_heads property used in LM Studio to glm4v_moe.LanguageModel by @hehua2008 in https://github.com/Blaizzy/mlx-vlm/pull/472
Fix: rope rotation (GLM-4.5v) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/481
Remove scipy dep by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/482
Add CUDA and CPU as optional deps by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/483
Fix Deepseek vl2 chat template by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/488
Fix deepseek-vl default chat template by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/490
Fix smolvlm video generate by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/491
Fix base64 encoded images by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/493
Map Apriel configs to pixtral and fix prompt formatting by @ivanfioravanti in https://github.com/Blaizzy/mlx-vlm/pull/518
Fix video understanding NB by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/519
Fix fine-tuning bug in trainer.py by @avishekjana in https://github.com/Blaizzy/mlx-vlm/pull/473
Bump minimum required Python version to 3.10 by @dokterbob in https://github.com/Blaizzy/mlx-vlm/pull/485
Add Qwen3-VL (Dense & MoE) by @Blaizzy & @vincentamato in https://github.com/Blaizzy/mlx-vlm/pull/528
Fix video Qwen3-VL by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/529
Fix Qwen3 VL (Dense) Sanitize by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/531
Bump version to 0.3.4 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/535

New Contributors

@hehua2008 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/472
@ivanfioravanti made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/518
@avishekjana made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/473
@dokterbob made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/485
@vincentamato made their first contributuin in https://github.com/Blaizzy/mlx-vlm/pull/529

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.3...v0.3.4

Open source →

0.3.320 Aug 2025

fix changelog task by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/441

What's Changed

fix changelog task by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/441
Add LFM2-VL by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/460
[llava_next] Fix config inheritance by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/448
Kimi_VL: Fix activation args by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/465
Add GLM-4-5V by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/458
External access to LFM2-VL merge-input-IDs method by @christian-lms in https://github.com/Blaizzy/mlx-vlm/pull/466
Add Command-A-Vision by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/467
[kernels] Use a header for bicubic_interpolate for compatibility with macOS < 15 by @mattjcly in https://github.com/Blaizzy/mlx-vlm/pull/469

New Contributors

@christian-lms made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/466

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.2...v0.3.3

Open source →

0.3.222 Jul 2025

Fix energy calc in omni.py by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/427

What's Changed

Fix energy calc in omni.py by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/427
Feat(build): migrate to pyproject.toml by @SauravMaheshkar in https://github.com/Blaizzy/mlx-vlm/pull/282
Fix quant predicate by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/430
Load dependencies from requirements.txt by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/431
Fix broken wheel builds caused by ambiguous package spec by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/432
Make UI and audio dependencies optional by @zhnext in https://github.com/Blaizzy/mlx-vlm/pull/433
Cleanup: Refactor config by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/437
Add cuda support by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/438
Fix/server module exposure by @zhnext in https://github.com/Blaizzy/mlx-vlm/pull/434
Support text only training by @Goekdeniz-Guelmez in https://github.com/Blaizzy/mlx-vlm/pull/424
Fix phi3_v and molmo mask by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/440

New Contributors

@zhnext made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/433
@Goekdeniz-Guelmez made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/424

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.1...v0.3.2

Open source →

0.3.112 Jul 2025

fix(chat-ui): Fix imports, blocking Chat-ui server start by @zenyr in https://github.com/Blaizzy/mlx-vlm/pull/420

What's Changed

fix(chat-ui): Fix imports, blocking Chat-ui server start by @zenyr in https://github.com/Blaizzy/mlx-vlm/pull/420
Fix server empty image for Gemma3n by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/422
[Gemma3n] Add hooks for image embedding computation by @will-lms in https://github.com/Blaizzy/mlx-vlm/pull/407
[gemma3n] Fix OCR after weight re-upload by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/425
Add gemma3n omni example by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/426

New Contributors

@zenyr made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/420

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.3.0...v0.3.1

Open source →

0.3.05 Jul 2025

[gemma3n] Correctly scale text embeddings for quantized gemma3n conversions by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/397

What's Changed

[gemma3n] Correctly scale text embeddings for quantized gemma3n conversions by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/397
smolvlm_video example: fix typo in system prompt by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/389
Fix gemma3n pixel casting by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/398
Fix audio model check and prompt utils by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/395
Add KV Quantization by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/401
Fix Gemma3n multi-task merging and update LM by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/405
[gemma3n] Fix vision encoder implementation of EdgeResidual and UniversalInvertedResidual by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/410
fix: Remove unnecessary unicode_escape decoding for Chinese text input by @nicekate in https://github.com/Blaizzy/mlx-vlm/pull/403
Add support for Mixed Quant by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/413
Fix gemma3n Vision OCR + LM only reponses by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/414
Fix generate signature by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/416
Add support for audio modality in server by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/417
Update server, readme and misc by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/418

New Contributors

@nicekate made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/403

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.2.0...v0.3.0

Open source →

0.2.026 Jun 2025

Fix Gemma 3 config by @DePasqualeOrg in https://github.com/Blaizzy/mlx-vlm/pull/384

What's Changed

Fix Gemma 3 config by @DePasqualeOrg in https://github.com/Blaizzy/mlx-vlm/pull/384
Add option to save LoRA adapter after each training epoch by @breynolds007 in https://github.com/Blaizzy/mlx-vlm/pull/385
De-duplicate Mixtral3's merge_input_ids_with_image_features by @will-lms in https://github.com/Blaizzy/mlx-vlm/pull/390
Add support for Gemma3n (Omni) by @FL33TW00D @pcuenca and @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/391
Add support for audio by @Blaizzy https://github.com/Blaizzy/mlx-vlm/pull/391

New Contributors

@DePasqualeOrg made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/384
@will-lms made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/390

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.27...v0.2.0

Open source →

0.1.278 Jun 2025

Fix README POST request typo by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/358

What's Changed

Fix README POST request typo by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/358
Fix add_eos_token_ids early exit by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/356
Add max_position_embeddings to TextConfig dataclasses by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/360
Fix revision argument not passed in load by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/361
Add initial MkDocs setup by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/362
Update deploy-docs.yml by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/363
Update update-changelog.yml by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/364
Update mkdocs configuration by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/365
Remove hardcoded pytorch install error message by @mattjcly in https://github.com/Blaizzy/mlx-vlm/pull/369
Update processor tests by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/370
Fix Kimi-VL Chat template by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/376
Fix multi images understand for Qwen2 and Qwen2.5 VL by @qnguyen3 in https://github.com/Blaizzy/mlx-vlm/pull/377
Qwen2.5 vl fix by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/378
Bump mlx to v0.26.0 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/381
Bump by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/382

New Contributors

@prncvrm made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/319
@qnguyen3 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/377

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.26...v0.1.27

Open source →

0.1.2613 May 2025

When converting models, default dtype to config.json's torch_dtype by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/323

What's Changed

When converting models, default dtype to config.json's torch_dtype by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/323
FastAPI server by @pappitti in https://github.com/Blaizzy/mlx-vlm/pull/321
feat: Resume training from saved adapter files by @breynolds007 in https://github.com/Blaizzy/mlx-vlm/pull/327
docs: Fix command-line params in LORA.md to use dash separators by @breynolds007 in https://github.com/Blaizzy/mlx-vlm/pull/331
Fix internVL3 Attribute Error by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/333
fix models/internvl_chat/vision.py transpose issue by @e1732a364fed in https://github.com/Blaizzy/mlx-vlm/pull/340
[gemma3] Improve generation quality when using multiple images by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/342
Update utils.py by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/322
Remove duplicate get_model_path by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/343
[phi3_v] Remove logit dtype conversion from float32 to uint32 by @neilmehta24 in https://github.com/Blaizzy/mlx-vlm/pull/329
Make Gemma3 and Qwen2.5VL text-image input merging functions public static by @mattjcly in https://github.com/Blaizzy/mlx-vlm/pull/335
Fix: remove self reference in static method prepare_inputs_for_multimodal by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/348
Fix VLMs upcasting by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/350
Bump version by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/351

New Contributors

@pappitti made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/321
@breynolds007 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/327
@e1732a364fed made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/340
@mattjcly made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/335

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.25...v0.1.26

Open source →

0.1.2521 Apr 2025

Fix snapshot download by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/318

What's Changed

Fix snapshot download by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/318

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.24...v0.1.25

Open source →

0.1.2420 Apr 2025

Deps: Remove Scipy MLX by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/301

What's Changed

Deps: Remove Scipy MLX by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/301
Add Stopping criteria by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/307
Migrate to mlx-lm switch layers by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/308
Refactor stop criteria by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/310
Migrate cache co-routine by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/312
Refactor gemma3 clip by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/313
Skip special tokens by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/316

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.23...v0.1.24

Open source →

0.1.2317 Apr 2025

Add internVL3 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/295

What's Changed

Add internVL3 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/295
Add Kimi VL Thinking by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/299
Fix LM only responses by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/296
Fix Qwen 2.x VL video understanding by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/300

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.22...v0.1.23

Open source →

0.1.2212 Apr 2025

chore: fix sliding mask by @FL33TW00D in https://github.com/Blaizzy/mlx-vlm/pull/272

What's Changed

chore: fix sliding mask by @FL33TW00D in https://github.com/Blaizzy/mlx-vlm/pull/272
chore: fix double bos by @FL33TW00D in https://github.com/Blaizzy/mlx-vlm/pull/269
docs: add note on PIL by @SauravMaheshkar in https://github.com/Blaizzy/mlx-vlm/pull/276
Fix example signatures by @zboyles in https://github.com/Blaizzy/mlx-vlm/pull/279
Gemma 3 - Clip fp16 activations by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/293
fixed mx.metal.get_peak_memory() depreciate by @vovw in https://github.com/Blaizzy/mlx-vlm/pull/290
Add LLaMA 4 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/287
Add Florence-2 examples by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/128

New Contributors

@FL33TW00D made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/272
@SauravMaheshkar made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/276
@zboyles made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/279
@vovw made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/290

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.21...v0.1.22

Open source →

0.1.2121 Mar 2025

fix remaining DEFAULT_TEMP by @omercelik in https://github.com/Blaizzy/mlx-vlm/pull/265

What's Changed

fix remaining DEFAULT_TEMP by @omercelik in https://github.com/Blaizzy/mlx-vlm/pull/265
Bump undefined default temp by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/267

New Contributors

@omercelik made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/265

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.20...v0.1.21

Open source →

0.1.2020 Mar 2025

Fix return tensors by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/262

What's Changed

Fix return tensors by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/262
bump version by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/263
Rename temp to temperature by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/264

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.19...v0.1.20

Open source →

0.1.1919 Mar 2025

Copy all json files by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/249

What's Changed

Copy all json files by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/249
Remove mask from _step by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/255
Add Mistral Small 3.1 by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/256

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.18...v0.1.19

Open source →

0.1.1818 Mar 2025

Fix Gemma 3 Text-only and prompt template by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/248

What's Changed

Fix Gemma 3 Text-only and prompt template by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/248

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.17...v0.1.18

Open source →

0.1.1712 Mar 2025

Add support for Gemma 3 by @pcuenca @FL33TW00D and @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/235

What's Changed

Add support for Gemma 3 by @pcuenca @FL33TW00D and @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/235

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.16...v0.1.17

Open source →

0.1.165 Mar 2025

Fix aya vision 32b by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/226

What's Changed

Fix aya vision 32b by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/226
Fix Qwen2-vl and Chat UI by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/227

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.15...v0.1.16

Open source →

0.1.155 Mar 2025

add status badge by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/215

What's Changed

add status badge by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/215
Fix Pixtral inputs by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/216
Add Computer Use with Voice Control Interface by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/217
Add stream_generate and GenerationResult to init.py by @asmeurer in https://github.com/Blaizzy/mlx-vlm/pull/202
Add Autonomous GUI Agents with Voice interface (Computer Use) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/220
Fix typo and local path by @lucataco in https://github.com/Blaizzy/mlx-vlm/pull/221
Add support for aya-vision by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/224

New Contributors

@asmeurer made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/202
@lucataco made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/221

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.14...v0.1.15

Open source →

0.1.1420 Feb 2025

Fix generate() call in object_detection.ipynb to follow the latest parameter order by @JoeJoe1313 in https://github.com/Blaizzy/mlx-vlm/pull/200

What's Changed

Fix generate() call in object_detection.ipynb to follow the latest parameter order by @JoeJoe1313 in https://github.com/Blaizzy/mlx-vlm/pull/200
Add support for smolvlm by @pcuenca in https://github.com/Blaizzy/mlx-vlm/pull/208

New Contributors

@JoeJoe1313 made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/200

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.13...v0.1.14

Open source →

0.1.131 Feb 2025

Add Video Understanding Support (beta) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/97

What's Changed

Add Video Understanding Support (beta) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/97
Add CLI chat by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/97

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.12...v0.1.13

Open source →

0.1.1229 Jan 2025

Add support for Qwen2-5-VL by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/189

What's Changed

Add support for Qwen2-5-VL by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/189
Fix idefics (2 and 3) do-image-split by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/191

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.11...v0.1.12

Open source →

0.1.1118 Jan 2025

Chat in CLI. by @chigkim in https://github.com/Blaizzy/mlx-vlm/pull/168

What's Changed

Chat in CLI. by @chigkim in https://github.com/Blaizzy/mlx-vlm/pull/168
Fix skip-vision predicate and add utils unit test (quantize and inputs) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/172
Refactor topk to use mlx.core (DS-VL2) by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/175
Fix trainer and Qwen2-VL by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/179
Pin latest mlx by @Blaizzy in https://github.com/Blaizzy/mlx-vlm/pull/184

New Contributors

@chigkim made their first contribution in https://github.com/Blaizzy/mlx-vlm/pull/168

Full Changelog: https://github.com/Blaizzy/mlx-vlm/compare/v0.1.10...v0.1.11

Open source →

Your coding agent can read these notes before it upgrades. Set up the MCP server →