Qwen3.8-27B (Lottolabs) runs on a Tenstorrent P150 32 GB via vLLM using BFP4/BFP8 mixed-precision quantization and DFlash2 speculative decoding.

Output speed across three measured runs: 56.4 tok/s fastest, 55.9 tok/s median, 23.8 tok/s slowest, with prefill at 1,212 tok/s on a 2,048-token prompt. The benchmark is flagged unverified due to missing prompt hash, missing output sample, and 206 output tokens below the 256-token minimum.