Tencent releases DSpine speculative decoding draft model for Qwen3.8-27B, beating DFlash2 on all benchmarks
teortaxesTex · x · 2026-11-08
Tencent open-sourced DSpine, a speculative decoding draft model for Qwen3.8-27B (5-layer, block size 16, 8192-token sliding window), claiming it beats the official DFlash2 across the board.
- How it works: DSpine generates draft tokens in parallel; each layer passes predicted predecessor features to adjacent positions, carrying causal dependencies through network depth.
- General benchmarks: acceptance length reaches 7.20 on GSM8K (vs 5.69 for DFlash2), 6.49 on MATH-500, 6.14 on MBPP, with greedy and temperature sampling both leading.
- Agent benchmarks: consistent gains of 0.5-1+ acceptance length on WorkBuddy Code/Web/Office, SWE-bench Pro, and τ²-bench, translating into meaningful end-to-end throughput speedups over autoregressive decoding.
- Runtime: SGLang 0.5.22, BF16, FlashAttention 3, dynamic verification with ReplaySSM.
Released under Apache-2.0 with two accompanying arXiv papers.
Related event: Tencent Open-Sources DSpine Draft Model, Speeding Up 27B Inference by 3x(4 posts)→
More from Infra
- Data-center water use debate: a post misunderstanding basic water science gets 22K likes — ATTlKA · 2026-11-10
- Qdrant presents vector index-native relevance feedback in production at CIKM Rome — qdrant_engine · 2026-11-10
- int8 quantization plus fp16 rescoring: 1.4M embeddings retrieved in under 2ms — tomaarsen · 2026-11-10
- int8 embeddings plus Usearch index: 1.4M vectors searched in under 2ms on CPU — tomaarsen · 2026-11-10
- Google rolls out broader AI chip lineup to challenge Nvidia in a market worth hundreds of billions — dinabass · 2026-11-10
- Thinking Machines engineer returns to Modal, reflecting on a year of frontier ML infrastructure work — josh_wills · 2026-11-10