# What changed in mlx-vlm

> 34 changes in mlx-vlm 0.7 from its official release notes, upgrade guides and advisories: 15 new APIs and options. Each one cited. Sources checked 2026-10-24. HTML: https://unstale.sh/pypi/mlx-vlm/

To see which of these apply to a repository, run `npx @unstale/cli init` in it.

## mlx-vlm 0.7

### New APIs and options

- Adds support for Muse Spark Open with audio, video, and images. (0.7.10) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.10)
- Adds native BF16 support for Flint. (0.7.10) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.10)
- Adds support for Flux-3 image generation and editing. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Adds GRPO support for vision-language models. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Adds live image resolution controls to the server. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Adds the MOSS-Image-SFT image generation model. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Adds support for Nemotron 3.5 Super VL with MTP and Bolt support. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Adds support for the Clef decision model. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Adds multimodal embedding support for EmbeddingGemma 2. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Adds support for the kolibri1 model. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Adds the LiquidAI d1-3B decision model. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Adds the LiquidAI d1-omni-600M decision model. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Adds MiniCPM-V 35B-A3B support with a custom processor. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Adds MiniCPM-V 4.7 1B support with a custom processor. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Adds support for the Xing4.0 model. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)

### Performance

- Batches offloaded MoE execution. (0.7.10) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.10)
- Reads strided K/V during ragged decode attention instead of copying the KV cache every step. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Releases the prompt cache when a SpeculativeGenerationBatch finishes. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)

### Behavior and default changes

- Simplifies vision cache handling for MiniCPM-V 4.7. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Tool arguments are streamed in Chat Completions and Responses. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)

### Fixes and other changes

- Fixes fused NVFP4 rowwise scales in expert offload repacking. (0.7.10) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.10)
- Preserves NVFP4 scales. (0.7.10) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.10)
- Fixes remaining affine QMV precision mismatches on MLX 0.32.4. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- The CLI and generation APIs now honor sampling defaults from checkpoints. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Reads PIL images as channels-last in EmbeddingGemma 2 processors. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Reads PIL images as channels-last in Gemma 4 image processors. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Fixes vision dtype handling with quantized embeddings. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Fixes image feature merging for Mistral Large 3. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Fixes rotating cache sinks and absolute position handling. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Uses PyTorch channel grouping in the pixel decoder GroupNorm. (0.7.9) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.9)
- Fixes compaction when maximum output fills the context window and allows compaction between completed tool groups. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Fixes cancellation during multi-row prefill. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Muse Glimmer now honors explicit RoPE scaling settings. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)
- Stabilizes tool-schema key order for prefix caching. (0.7.8) — [release notes](https://github.com/Blaizzy/mlx-vlm/releases/tag/v0.7.8)