Tencent releases DSpine draft model for Qwen3.8-27B
Original title: tencent/Qwen3.8-27B-DSpine-b16 — new model trending #30 on Hugging Face
The five-layer checkpoint uses block size 16 and an 8,192-token window and is trending at number 30.
Tencent published tencent/Qwen3.8-27B-DSpine-b16, a five-layer draft model that generates tokens in parallel for Qwen3.8-27B. The card reports higher acceptance lengths than DFlash2 on GSM8K, MATH-500, HumanEval, MBPP, MT-Bench, and several agent benchmarks under SGLang 0.5.22. Throughput figures versus autoregressive decoding are only partly shown.
Unconfirmed: Acceptance and speedup figures are the project's own measurements, and the throughput table is truncated.
- Draft block size is 16, compared with DFlash2 block size 8.
- GSM8K greedy acceptance length is 7.20 for DSpine versus 5.69 for DFlash2.
- Runtime listed is SGLang 0.5.22, BF16, FlashAttention 3, concurrency 8.
- Agent tests allow up to 32,768 new tokens per request.
- Organizations
- Tencent
- AI models
- Qwen3.8-27B-DSpine-b16Qwen3.8-27B
From the original article
4 min read at huggingface.cotext-generation · 1,464 downloads · 143 likes · license apache-2. 0
By tencent · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.