pypi / mlx-vlm

mlx-vlm

What changed in mlx-vlm 0.7, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-24.

15 new APIs and options3 performance2 behavior and defaults14 fixes and other

See which of these apply to your code, ranked, in your agent:

npx @unstale/cli init

mlx-vlm 0.734 changes

Adds support for Muse Spark Open with audio, video, and images.
0.7.10release notes
Adds native BF16 support for Flint.
0.7.10release notes
Adds support for Flux-3 image generation and editing.
0.7.9release notes
Adds GRPO support for vision-language models.
0.7.9release notes
Adds live image resolution controls to the server.
0.7.9release notes
Adds the MOSS-Image-SFT image generation model.
0.7.9release notes
Adds support for Nemotron 3.5 Super VL with MTP and Bolt support.
0.7.9release notes
Adds support for the Clef decision model.
0.7.8release notes
Adds multimodal embedding support for EmbeddingGemma 2.
0.7.8release notes
Adds support for the kolibri1 model.
0.7.8release notes
Adds the LiquidAI d1-3B decision model.
0.7.8release notes
Adds the LiquidAI d1-omni-600M decision model.
0.7.8release notes
Adds MiniCPM-V 35B-A3B support with a custom processor.
0.7.8release notes
Adds MiniCPM-V 4.7 1B support with a custom processor.
0.7.8release notes
Adds support for the Xing4.0 model.
0.7.8release notes

Batches offloaded MoE execution.
0.7.10release notes
Reads strided K/V during ragged decode attention instead of copying the KV cache every step.
0.7.9release notes
Releases the prompt cache when a SpeculativeGenerationBatch finishes.
0.7.9release notes

Simplifies vision cache handling for MiniCPM-V 4.7.
0.7.9release notes
Tool arguments are streamed in Chat Completions and Responses.
0.7.8release notes

Fixes fused NVFP4 rowwise scales in expert offload repacking.
0.7.10release notes
Preserves NVFP4 scales.
0.7.10release notes
Fixes remaining affine QMV precision mismatches on MLX 0.32.4.
0.7.9release notes
The CLI and generation APIs now honor sampling defaults from checkpoints.
0.7.9release notes
Reads PIL images as channels-last in EmbeddingGemma 2 processors.
0.7.9release notes
Reads PIL images as channels-last in Gemma 4 image processors.
0.7.9release notes
Fixes vision dtype handling with quantized embeddings.
0.7.9release notes
Fixes image feature merging for Mistral Large 3.
0.7.9release notes
Fixes rotating cache sinks and absolute position handling.
0.7.9release notes
Uses PyTorch channel grouping in the pixel decoder GroupNorm.
0.7.9release notes
Fixes compaction when maximum output fills the context window and allows compaction between completed tool groups.
0.7.8release notes
Fixes cancellation during multi-row prefill.
0.7.8release notes
Muse Glimmer now honors explicit RoPE scaling settings.
0.7.8release notes
Stabilizes tool-schema key order for prefix caching.
0.7.8release notes