148 likesNew · 3d old
Tencent's DSpine speeds up Qwen3.8 with parallel drafts
What is Qwen3.8-27B-DSpine-b16?
DSpine is a draft model that predicts multiple tokens at once, passing predecessor features across layers to keep causal order. It runs on Qwen3.8-27B with a 16-token block size and an 8,192-token sliding window.
Who it is for
If you run large language models and need higher throughput, DSpine can shave latency without losing quality.
The story so far
Early Signal first spotted tencent/Qwen3.8-27B-DSpine-b16 on November 9, 2026, with 148 likes. On its busiest day it gained 49 likes. The project is 3 days old. Early Signal files it under LLMs, Agents, Coding.
Questions about Qwen3.8-27B-DSpine-b16
- What do I need to run Qwen3.8-27B-DSpine-b16?
- GPU with BF16 and FlashAttention 3 support, Python environment with SGLang 0.5.22