This development release updates mlx-vlm to 0.6.1 (041f889) to add Gemma4 Unified (12B) support, adds cohere2_moe (Command A+) support, and includes several bug fixes.
- Updated
mlx-vlmto 0.6.1 (041f889) with Gemma4 Unified long-text prefill fixes. - Added support for Gemma4 Unified (12B) through the updated mlx-vlm backend.
- Added support for
cohere2_moemodels, including Command A+. - Fixed streaming detokenization for raw VLM/Gemma4 paths so byte-fallback tokenizers no longer emit replacement characters in split UTF-8 output.
- Added VLM MTP controls to the macOS model settings UI.
- Exposed advanced DFlash model settings in the macOS app, including the
ddtreeverify mode. - Fixed speculative-control state in the macOS app while VLM MTP is enabled.
- Restored the interactive Claude model picker in
omlx launch.