tencent/Qwen3.8-27B-DSpine-b16
AI model · https://huggingface.co/tencent/Qwen3.8-27B-DSpine-b16
First seen by Oriel on .
AI summary
The Hugging Face checkpoint is a five-layer DSpine draft model for Qwen3.8-27B, reporting higher acceptance lengths and up to 3.84× end-to-end throughput versus autoregressive decoding.
- DSpine generates draft tokens in parallel, passing predicted predecessor features to adjacent positions through network depth; this checkpoint uses block size 16 and an 8,192-token sliding window.
- On general benchmarks, DSpine acceptance length exceeded DFlash2 across GSM8K, MATH-500, HumanEval, MBPP, and MT-Bench at both temperatures.
- On agent benchmarks, DSpine reached 4.77–5.63 acceptance length and 2.71–3.84× autoregressive throughput, depending on task and reasoning setting.
- Usage requires DSpine with SGLang, BF16, FlashAttention 3, ReplaySSM, and the Qwen3.8-27B base model configured for speculative decoding.
- The checkpoint is released under the Apache 2.0 license, with installation provided through the Tencent/DSpine repository.
#speculative decoding #dspine #qwen #sglang
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Likes
Total likes reported by the source.
- Hugging Face Models: 285, latest observation ; +185 over observed period since
Downloads (30d)
Rolling 30-day downloads. A change in this value is not new downloads during the observed period.
- Hugging Face Models: 5,984, latest observation ; +5,332 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Models | tencent/Qwen3.8-27B-DSpine-b16 | https://huggingface.co/tencent/Qwen3.8-27B-DSpine-b16 Read discussion |