DSpine: A parallel draft acceleration model for Qwen3.8-27B by Tencent

A five-layer DSpine draft model trained by Tencent for Qwen3.8-27B, used for speculative decoding: it generates draft tokens in parallel, with each layer passing predicted preceding features to adjacent positions, relying on network depth to propagate causal dependencies. Block size 16, sliding window 8,192 tokens. Official tests on SGLang show higher average acceptance lengths across various tasks compared to DFlash2; in agent replays on WorkBuddy code/web/office, SWE-bench Pro, and τ²-bench, end-to-end throughput is approximately 2.7–3.8 times that of standard autoregressive decoding.

Original title: tencent/Qwen3.8-27B-DSpine-b16

A five-layer DSpine draft model trained by Tencent for Qwen3.8-27B, used for speculative decoding: it generates draft tokens in parallel, with each layer passing predicted preceding features to adjacent positions, relying on network depth to propagate causal dependencies. Block size 16, sliding window 8,192 tokens. Official tests on SGLang show higher average acceptance lengths across various tasks compared to DFlash2; in agent replays on WorkBuddy code/web/office, SWE-bench Pro, and τ²-bench, end-to-end throughput is approximately 2.7–3.8 times that of standard autoregressive decoding.

Why it matters Increases inference throughput without changing the main model, suitable for deploying Qwen3.8-27B with SGLang for agent and code tasks

Rising score 146 · Likes 148 · Downloads 1464

Reasons

  • Specifically tested acceptance length and throughput in multi-turn agent dialogue scenarios, not just math and coding problems
  • Includes paper and GitHub code, Apache-2.0

Usage

  • After installing Tencent/DSpine, add --speculative-algorithm DSPINE to sglang serve, point --speculative-draft-model-path to this model, and follow the model card for other parameters

Risks

  • Only compatible with the Qwen3.8-27B main model, depends on SGLang and specific backend configurations, high barrier for personal GPU deployment
  • Acceleration data was measured by the official team under specific concurrency and configurations
  • This entry has not been run; refer to the model card for actual performance

Open

Hugging Face Trending

Explore the latest AI Radar updates