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