Tencent open-sourced DSpine, a parallel drafter for Qwen3-8B speculative decoding
Tencent released Qwen3-8B-DSpine-b16 on Hugging Face under Apache 2.0. The model is a parallel drafter for speculative decoding, achieving an average acceptance length of 7.83 at temperature 0 on Qwen3-8B. Inference code was released in October 2026.
China context
- Original name
- 腾讯
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the weights from Hugging Face and integrate DSpine as a drafter for Qwen3-8B speculative decoding under Apache 2.0.
- For investors
- Tencent's release of a high-acceptance drafter under Apache 2.0 may pressure commercial speculative decoding solutions, as it lowers the cost barrier for efficient Qwen3-8B inference.
Tencent published Qwen3-8B-DSpine-b16 on Hugging Face under an Apache 2.0 license. DSpine is a parallel drafter for speculative decoding where each position receives the predicted feature of its predecessor, allowing the causal chain to unfold over network depth while all positions are drafted in parallel. The model card reports acceptance lengths on Qwen3-8B, with DSpine achieving 7.83 average at temperature 0 and 6.98 at temperature 1, outperforming DFlash, Domino, DSpark, and DFlash2. Inference code was released in October 2026, with training code and a 27B version still on the roadmap.
DSpine uses adjacent causal injection: at every draft layer, each position receives the predicted feature of its predecessor, so the causal chain unfolds over network depth while all positions are drafted in parallel. Reported acceptance lengths on Qwen3-8B are 7.83 (T=0) and 6.98 (T=1), higher than DFlash, Domino, DSpark, and DFlash2. The model is retrained on Open-PerfectBlend with the same configuration as DSpark and DFlash2.
Developers using speculative decoding can now test a higher-acceptance drafter for Qwen3-8B under Apache 2.0, potentially reducing serving cost per token. Competitors offering speculative decoding frameworks must match or exceed DSpine's reported acceptance lengths to remain attractive.
Apache 2.0 licensing allows commercial use, and higher acceptance lengths can lower inference costs for Qwen3-8B deployments. However, the reported numbers are company-reported and not independently verified.
The roadmap lists training code for DSpine and a DSpine for Qwen3.8-27B as not yet released. A verifiable next signal is whether Tencent publishes the training code or the 27B checkpoint on Hugging Face.