qwenfast is a high-throughput Qwen3.8-27B inference stack using speculative decoding and a custom Triton DeltaNet kernel, hitting 133 t/s on H200.