Tencent releases Qwen3.8-27B DSpine draft checkpoint
Tencent published the Apache 2.0-licensed Qwen3.8-27B-DSpine-b16 checkpoint on Hugging Face, a five-layer, 2B-parameter draft model with a block size of 16 and an 8,192-token sliding window. Its model card says DSpine generates draft tokens in parallel and reports higher acceptance lengths and throughput than DFlash2, with end-to-end results reaching 3.84× autoregressive decoding across the listed agent benchmarks.
The checkpoint provides a concrete implementation of parallel speculative decoding for Qwen3.8-27B, while its reported throughput gains offer a relevant new option for model inference.
tencent/Qwen3.8-27B-DSpine-b16 Text Generation • 2B • Updated 5 days ago • 6.92k • 316
license: apache-2.0 base_model:
- Qwen/Qwen3.8-27B pipeline_tag: text-generation inference: false tags:
- dspine
- speculative-decoding
- draft-model
- sglang
DSpine for Qwen3.8-27B
DSpine generates draft tokens in parallel. Each layer passes predicted predecessor features to adjacent positions, carrying causal dependencies through network depth.
#Tencent#Qwen3.8-27B#Qwen3.8-27B-DSpine-b16#DSpine#SGLang#DFlash2#DSpark