What is Qwen3.8-27B DSpine?
DSpine is not a model you chat with. It is a small five-layer draft model that sits beside Qwen3.8-27B and proposes 16 tokens at a time, which the full model then checks in one pass; accepted tokens come out much faster than generating one by one. Tencent's method drafts all 16 positions in parallel while passing each position's predicted features to the next, which keeps the drafts consistent.
On Tencent's agent benchmarks, served with SGLang on BF16 at eight concurrent requests, it delivered 2.71 to 3.84 times the output speed of plain decoding, ahead of the official DFlash2 draft model at 2.33 to 3.17 times. It is released under Apache 2.0 with a research paper on arXiv.
Is Qwen3.8-27B DSpine worth it?
Speculative decoding is the rare optimisation that costs nothing in quality: a small draft model guesses several tokens, the large model verifies them in one pass, and only tokens the large model would have produced are kept. DSpine is Tencent's draft model for Qwen3.8-27B, five layers deep, proposing blocks of 16 tokens.
The gains on the model card are large and measured on realistic work. On WorkBuddy agent tasks for code, web and office, SWE-bench Pro and τ²-bench, served with SGLang 0.5.22 at concurrency 8, DSpine lifted throughput 2.71 to 3.84 times over plain decoding with greedy sampling, against 2.33 to 3.17 times for the official DFlash2 drafter. Average accepted length rose from about 4 to 5 tokens per step to about 5 to 7.
The constraints are practical. It works with SGLang and the DSpine package only, needs FlashAttention 3, and the reference command assumes a long list of flags tuned for Qwen3.8's hybrid attention. Gains shrink at higher concurrency, where GPUs are already busy, and the benchmarks are Tencent's own, run at eight concurrent requests. Because the arXiv paper is a preprint, the method has not been peer reviewed.
Where it earns its place
A start-up running a Qwen3.8-27B coding agent on two H100s
Cut the wait on each agent step
Add DSpine as the draft model in SGLang with 16 draft tokens; the model card reports about 2.7 to 3.2 times faster code tasks. The weights are free, and the same GPUs serve more requests per hour.
An ML engineer comparing draft models
Pick between DFlash2 and DSpine for production
Run both on your own traffic at your real concurrency, since Tencent's numbers come from eight concurrent requests and gains fall as load rises.
Before you commit
- Setup
- Clone Tencent/DSpine and pip install it, download the draft weights, then start sglang serve with --speculative-algorithm DSPINE and the flags on the model card.
- Integrations
- SGLang 0.5.22 with FlashAttention 3 and FlashInfer; Qwen3.8-27B as the target model.
- Data and privacy
- Runs on your own servers; weights under Apache 2.0.
- Cost at scale
- No licence cost; faster decoding lets the same GPUs handle more requests, though the draft model takes a little extra memory.
- Limits
- Tied to Qwen3.8-27B and SGLang; gains fall at high concurrency; preprint method.
What’s good
- About 2.7 to 3.8 times faster on agent tasks, per the model card
- Output matches the base model exactly
- Beats the official DFlash2 drafter on every listed task
- Apache 2.0 open weights
What’s not
- SGLang and the DSpine package only
- Benchmarks run by Tencent at concurrency 8
- Gains shrink when the GPU is already saturated
- Only useful with Qwen3.8-27B
Best for
Teams serving Qwen3.8-27B with SGLang who want lower latency per request without changing the model's answers.
Qwen3.8-27B DSpine: common questions
- Is Qwen3.8-27B DSpine worth it?
- A free speed-up for anyone serving Qwen3.8-27B: roughly three times faster generation on agent workloads, with output identical to the base model, if you run SGLang. We rate it 4.2 out of 5.
- How much does Qwen3.8-27B DSpine cost?
- Free open weights under Apache 2.0; you pay only for the GPUs that serve Qwen3.8-27B.
- What are the downsides of Qwen3.8-27B DSpine?
- SGLang and the DSpine package only; benchmarks run by Tencent at concurrency 8; gains shrink when the GPU is already saturated; only useful with Qwen3.8-27B.
- Who is Qwen3.8-27B DSpine best for?
- Teams serving Qwen3.8-27B with SGLang who want lower latency per request without changing the model's answers.
- How much does Qwen3.8-27B DSpine cost for a team or in production?
- No licence cost; faster decoding lets the same GPUs handle more requests, though the draft model takes a little extra memory.
- Is Qwen3.8-27B DSpine safe for company data?
- Runs on your own servers; weights under Apache 2.0.
Answers checked Nov 15, 2026.