Tencent has released DSpine, an open-source draft model that speeds up Qwen3.8-27B inference. It drafts tokens in parallel and achieves up to a 3.84 times speedup across complex agent benchmarks.
FireBrief
FireBrief