v0.6.16

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