TensorFold doubles Qwen 27B decode to 158 tok/s on M3 Ultra with parallel-lane speculative decoding

TheMoonMidas · x · 2026-09-27

TensorFold, an open-source inference engine for Apple Silicon built on MLX + Metal, released v0.3.4 enabling parallel lanes on M1–M4 chips. It more than doubled Qwen3.8-27B code decode on M3 Ultra to 141–158 tok/s using a "draft wide, commit exact" scheme: candidate tokens are drafted in parallel and verified against the target model in one pass. Reported measurements include 188–206 tok/s for Nemotron 3.5 Lightning 30B-A3B on M5 Max, and 120–124 tok/s for Qwen3.8 27B with DFlash2. Ships with an OpenAI-compatible API; lane batching for concurrency is next.

Original post →

More from Infra

Infra channel →