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