Tencent releases a DSpine draft model for Qwen3.8-27B

Qwen3.8-27B-DSpine-b16 is an approximately two-billion-parameter draft model for speculative decoding. Its Apache 2.0 release includes instructions for running it with SGLang. In Tencent’s GSM8K evaluation at temperature zero, average acceptance length rose from 5.69 with DFlash2 to 7.20.

Visit original source