Tencent Open-Sources DSpine Draft Model, Speeding Up 27B Inference by 3x

Tencent released and open-sourced DSpine, a draft model method for parallel speculative decoding, along with a Qwen3.8-27B draft model on Hugging Face. It achieves up to 27.8% higher acceptance length over DFlash2 and reportedly speeds up 27B inference by 3x.

2026-11-07 ~ 2026-11-09 · 4 related posts

• Tencent releases Qwen3.8-27B-DSpine: a 27B draft model for speculative decoding — tencent · 2026-11-07
• Tencent's DSpine lifts speculative decoding acceptance length 27.8% over DFlash — pmttyji · 2026-11-07
• Tencent releases DSpine speculative decoding draft model for Qwen3.8-27B, beating DFlash2 on all benchmarks — teortaxesTex · 2026-11-08
• Tencent's New Speculative Decoding Method Claims 3x Faster 27B Inference — emdeka87 · 2026-11-09