Tencent's DSpine Speeds Up Qwen3.8-27B Inference 3x

Tencent's DSpine speculative decoding uses a 5-layer draft model that predicts 16 tokens in one forward pass, boosting Qwen3.8-27B inference up to 3x over the 1.6x of traditional MTP. The draft model is on HuggingFace and has already been ported to llama.cpp with GGUF quants.

2026-11-09 ~ 2026-11-10 · 3 related posts

• DSpine boosts Qwen3.8-27B to 94-132 tok/s in llama.cpp, 1.75x over DFlash2 — pbaylies · 2026-11-09
• Tencent's DSpine draft model lifts Qwen3.8-27B speculative decoding to 3x, 805 tps — karminski3 · 2026-11-10
• DSpine explained: 5-layer draft model guesses 16 tokens in one forward pass — karminski3 · 2026-11-10