you ‧ notes ‧ unread ‧ starred ‧ add url

@TeksEdge: 🍎⚡ Qwen3.8-27B is now hitting ~70 tok/s on an M5 Max MacBook Pro with MLX. And t

https://x.com/TeksEdge/status/2090249962840043832

august 2026 ‧ twitter twitter-post imported ‧ open in karakeep

by David Hendrickson ‧ X (formerly Twitter) ‧ published 2026.08.20 ‧ saved copy from 2026.08.28

From X bookmark — @TeksEdge (David Hendrickson), Aug 20 2026 https://x.com/TeksEdge/status/2090249962840043832

David Hendrickson announced that Qwen3.8-27B now achieves around 70 tokens per second on an M5 Max MacBook Pro using MLX, without sacrificing model size or output quality. This speedup is enabled by DFlash 2, which leverages speculative decoding to generate multiple candidate tokens in parallel while verifying them. Users can run this locally with two models and the updated oMLX software, avoiding cloud inference or specialized GPUs. However, some divergence in longer outputs compared to plain autoregressive decoding was noted.

regenerate summary

page description:

🍎⚡ Qwen3.8-27B is now hitting ~70 tok/s on an M5 Max MacBook Pro with MLX. And this isn’t a smaller model or a quality-reducing speed hack. DFlash 2 landed for Qwen3.8-27B, using speculative decoding to generate multiple candidate tokens in parallel while Qwen verifies them.

« back to bookmarks