Qwen3.8-27B local test hits 159 tok/s via speculative decoding

On October 9, 2026, benchmark data on running Qwen3.8-27B locally surfaced on Reddit, and a developer later published a write-up on Juejin breaking down the setup. According to that account, the tester combined Q8 DFlash2 speculative decoding with LemonSeed Engine 0.5.8 to push decoding speed from 14 tok/s to 159 tok/s.

The reported figures are 159.4 tok/s on a Sapphire Radeon AI PRO R9700 and 64.4 tok/s on the Strix Halo mobile platform. For comparison, the article cites a Mac Studio with M3 Ultra running Qwen3.8-27B Q4_K_M through Ollama, where the weights take up about 17 GB and sustained generation sits at roughly 14 tok/s.

The write-up attributes the speedup to two layers of optimization: draft tree verification and fusion of the gate/up GEMM operators. It also mentions comparing hardware and software configurations alongside llm-bench quality results, though the summary does not include the specific outcomes of those comparisons. The public information currently stops at this benchmark data and the breakdown of the optimization path, with no further developments reported in the available material.