Tencent DSpine: a speculative-decoding draft model that runs Qwen3.8-27B about 3x faster

Speculative decoding speeds up generation by letting a small draft model guess several upcoming tokens, which the large model then checks in a single pass. DSpine is Tencent's drafting method: it guesses 16 tokens at once, and each layer passes predicted features to neighboring positions so the guesses still account for token order. The new release is a five-layer draft model for Qwen3.8-27B under Apache 2.0. Install the SGLang plugin from the repo, then point the server at the draft model.

git clone https://github.com/tencent/dspine.git && cd dspine && python -m pip install -e .
sglang serve --model-path /path/to/Qwen3.8-27B --speculative-algorithm DSPINE --speculative-draft-model-path tencent/Qwen3.8-27B-DSpine-b16 --speculative-num-draft-tokens 16 --trust-remote-code

The repo README lists the other flags Qwen3.8-27B needs, including ReplaySSM and dynamic verification tuning.

Many teams run Qwen3.8-27B on a single GPU as a coding agent, so faster responses on the same hardware are easy to notice. Replaying coding, web and office agent conversations plus SWE-bench Pro, Tencent measured 2.7 to 3.8 times the throughput of plain decodingVendor claim, with more tokens accepted per step than the existing open DFlash2 drafterVendor claim. Output quality stays the same and only speed changes, so it's worth a test if you're trying to cut serving costs.