Qwen3.8-27B on 2x 3090 + vLLM + DFlash2: 218 tok/s single request

Collecting trend data

The percentage is based on available heat signal, not comment count or independent people.