Tencent releases DSpine draft model for Qwen3.8-27B: 16-token blocks, up to 3.84× agent speedup over autoregressive deco

Tencent has released Qwen3.8-27B-DSpine-b16, a five-layer draft model for speculative decoding with Qwen3.8-27B. It uses a block size of 16 and an 8,192-token sliding window. The checkpoint is listed at 2B parameters in BF16 Safetensors and is licensed under Apache 2.0.

How DSpine works

According to the model card, DSpine generates draft tokens in parallel. Each layer passes predicted predecessor features to adjacent positions, which carries causal dependencies through network depth. The accompanying paper is titled "Draft in Parallel, Condition Through Depth: Adjacent Causal Injection for Speculative Decoding" (arXiv:2609.36173, published Sep 28, authors Haohui Zhang, Keyu Chen, Haocheng Sun, Weibo Gu, Ruizhi Qiao, Xing Sun, Bo Jiang).

Evaluation setup

Tencent compares DSpine against the official DFlash2 draft model. The runtime is SGLang 0.5.22 in BF16 with FlashAttention 3 at concurrency 8. DSpine uses a 16-token draft block and DFlash2 uses 8. DSpine inference uses dynamic verification referencing DSpark, with ReplaySSM enabled.

General benchmarks use xhigh reasoning, greedy and temperature sampling, and at most 4,096 new tokens. Temperature runs use temperature=1.0, top_p=0.95, top_k=20.

Agent benchmarks cover WorkBuddy Bench (Code, Web, Office) using CodeBuddy sandbox trajectories, plus SWE-bench Pro and τ²-bench. They run at low and xhigh reasoning with temperature sampling, 20 conversations per subset, all tool-call prefixes replayed in order, prefix caching, and up to 32,768 new tokens per request.

General acceptance length (DFlash2 vs. DSpine)

Task	T=0	T=1
GSM8K	5.69 vs. 7.20	5.45 vs. 6.31
MATH-500	5.49 vs. 6.49	5.27 vs. 5.66
HumanEval	4.87 vs. 5.71	4.34 vs. 4.71
MBPP	5.04 vs. 6.14	4.79 vs. 5.41
MT-Bench	4.11 vs. 4.48	3.89 vs. 4.03

Agent throughput speedup over autoregressive decoding

Throughput is total output tokens divided by end-to-end concurrent replay wall time. The table shows the speedup over autoregressive (AR) decoding that the model card reports.

Benchmark	Low: DFlash2	Low: DSpine	xhigh: DFlash2	xhigh: DSpine
WorkBuddy Code	2.33×	2.71×	2.85×	3.18×
WorkBuddy Web	2.59×	3.13×	2.91×	3.33×
WorkBuddy Office	3.17×	3.84×	3.08×	3.55×
SWE-bench Pro	2.72×	2.95×	2.87×	3.19×
τ²-bench	2.96×	3.20×	2.56×	2.87×

The model card also reports agent acceptance lengths for each benchmark at both reasoning levels, with DSpine ahead of DFlash2 in every row.

Deployment

Installation requires cloning the Tencent/DSpine GitHub repository and running pip install -e .. Serving uses SGLang with --speculative-algorithm DSPINE and --speculative-num-draft-tokens 16. The published command also sets BF16, the FA3 attention backend, --enable-linear-replayssm-spec, --context-length 131072, --max-running-requests 8, the qwen3 reasoning parser and the qwen3_coder tool-call parser. No Inference Provider currently deploys the model. The page shows 1,464 downloads last month, one finetune and four quantizations of the checkpoint. The authors thank DFlash, SpecForge, SGLang and FlashInfer.

What this means

The reported gains compare DSpine against DFlash2 under the same runtime, precision and concurrency, so they speak to draft quality and not to a change of serving stack. DSpine's acceptance length is higher in every general and agent row listed, and its speedup over AR is higher in every agent row. All results come from Tencent's own model card at concurrency 8, and the page gives no independent replication. Teams should expect to use the DSpine repository and the specific SGLang flags above, since the speculative algorithm and ReplaySSM options are part of the published recipe.