v26.8.11-pre-release.1 — Qwen 3.8 Flash Next, MLX 0.32.2

Highlights

• Qwen 3.8 Flash Next runs natively. The 125B-A6B hybrid with its 51B n-gram table and 4B speculative head: hyper-connection residual streams, n-gram embeddings gathered from disk, sparse attention past 2k tokens, all on custom kernels. About 60 tok/s serial on an M4 Max, 78 tok/s with MTP, ~75 GB resident for the 4-bit pack (ddalcu/Qwen3.8-Flash-Next-MLX-Serve-4bit).
• It sees images and video too. The checkpoint's own vision tower is in the pack, with M-RoPE through both attention and the sparse-attention indexer. Follow-up questions about the same picture reuse the prefix cache instead of re-encoding it.
• Long prompts on Flash Next stay on the fast path. Sparse attention now has fused kernels for prefill and for multi-token verify, so MTP at an 8.5k-token prompt no longer loses to serial decode (was -28%).
• MTP on Flash Next is opt-in (--mtp or "enable_mtp": true): +41% on code, a few percent slower on prose, so you pick.
• MLX 0.32.2. Up to +3% decode on MoE models at long context.
• Fix: video input on Qwen 3.5/3.8 never reached the model. The frames were encoded but never spliced into the prompt, so the model described a clip it had not seen.

WARNING ⚠️

When you pull Qwen 3.8 Flash Next via the App, it doesnt pull the ngram bin file, but you can use HF CLI or mlx-serve pull CLI, to get it instead. Will have this fixed shortly.