jundot/omlx v0.2.14

on GitHub

Download the DMG that matches your macOS version (sequoia or tahoe).
If you're on an M5 Mac, you must use the macos26-tahoe DMG for M5 Neural Accelerator.

New Models

• Qwen3-VL embedding and reranking support (via mlx-embeddings 6e2ef52)
• Moondream3 vision-language model support (via mlx-vlm b7f853a)

Bug Fixes

• fix respect per-model max_tokens settings (#258)
• fix download stall timeout too short (120s → 300s) (#254)
• fix Qwen3.5 batch dimension mismatches under continuous batching (upstream mlx-vlm db3d558)
• fix Qwen3-VL attention mask slicing with mx.array kv_seq_len (upstream mlx-vlm)
• fix Qwen3-Omni integration (upstream mlx-vlm b7f853a)

Dependency Updates

• mlx >=0.29.2 → >=0.31.1
• mlx-embeddings 88522e2 → 6e2ef52
• mlx-vlm 348466f (0.3.13) → b7f853a (0.4.0)

full changelog: v0.2.13...v0.2.14