What is DSpine?
DSpine is a speculative-decoding method and SGLang plugin from Tencent, published with a paper in September 2026 and code from October. Speculative decoding uses a small draft model to guess several tokens ahead, which the main model then checks in one pass, so generation is faster without changing what the main model would have written. DSpine's draft proposes a block of 16 tokens at once and passes each position's predicted feature to the next at every layer, so guesses later in the block stay consistent with earlier ones.
There are draft weights for two target models: Qwen3-8B, and since 7 November Qwen3.8-27B, a five-layer draft of about 2 billion parameters with an 8,192-token sliding window. The code runs as an SGLang 0.5.22 plugin on CUDA, with an Apple silicon path through mlx-lm for Qwen3-8B only; training code sits in SpecForge.
Code and weights are under Apache 2.0 (GitHub shows the licence as unrecognised because of a Tencent preamble in the file). The repository has about 234 stars and no tagged releases.
Is DSpine worth it?
DSpine's case rests on Tencent's agent replays. Serving Qwen3.8-27B at eight concurrent requests, it reports end-to-end throughput 2.71 to 3.84 times that of plain decoding across WorkBuddy coding, web and office traces, SWE-bench Pro and τ²-bench. z-lab's official DFlash2 drafter for this model managed 2.33 to 3.17 times on the same replays. On WorkBuddy Office with low reasoning, that is 805.53 tokens a second against 664.77 for DFlash2 and 210.03 without speculation.
The gain comes from longer accepted guesses. The model card reports an average of 5.63 tokens accepted per step on WorkBuddy Office against 4.24 for DFlash2, and the paper, on Qwen3-8B, raises the mean acceptance length from DFlash's 3.77 to 4.82 and serving throughput by 23.3% over DFlash. All of these are Tencent's numbers, and the card does not name the GPU it used, though the FlashAttention 3 backend in its command implies a Hopper-class card.
The catch is reach. A drafter is trained for one target model, and DSpine has two. The 27B path needs a long list of SGLang flags, ReplaySSM for Qwen3.8's linear-attention layers and per-workload tuning of its dynamic verification. The MLX build covers only Qwen3-8B, and the CUDA plugin is pinned to a specific SGLang version.
Our take: speculative decoding is the cheapest speed-up a serving team can get, because it changes nothing about the answers. If Qwen3.8-27B is already your agent model on SGLang, DSpine is worth an afternoon of benchmarking against DFlash2 on your own traffic; otherwise wait for a drafter trained for your model.
Where it earns its place
A platform engineer running Qwen3.8-27B as the model behind an internal coding agent
Serve more agent sessions per GPU without changing answers
Add the 2-billion-parameter draft to the existing SGLang server with --speculative-algorithm DSPINE; Tencent reports about 3 times plain-decoding throughput on SWE-bench Pro replays. Run the SPS tuning script on your own traces first.
A developer running Qwen3-8B locally on an M-series Mac
Speed up single-user generation without a GPU server
Install the dspine_mlx package and run dspine-generate with an unmodified mlx-lm target; Tencent says it needs about 21 GB of unified memory for a BF16 Qwen3-8B target.
An inference researcher comparing parallel drafters
Reproduce DSpine against DFlash2 and DSpark on a new model
Training code is in SpecForge pull request 1038, trained on the Open-PerfectBlend data in Tencent's runs. The paper reports a 27.8% longer mean acceptance than DFlash on Qwen3-8B at temperature zero.
Before you commit
- Setup
- git clone the repository and pip install -e . into an SGLang 0.5.22 environment with Python 3.11+ and CUDA, then start sglang serve with the DSPINE flags from the model card; the draft downloads from Hugging Face.
- Integrations
- SGLang plugin (CUDA) and an SGLang-MLX plugin for Apple silicon; drafts trained with SpecForge; uses DSpark-style dynamic verification and ReplaySSM for Qwen3.8's linear-attention layers.
- Data and privacy
- Runs entirely on your hardware with no telemetry described. Apache 2.0 for code and weights, with third-party components under their own licences per THIRD_PARTY_NOTICES.
- Cost at scale
- Free software; the cost is the GPU already running the target model plus memory for a 2-billion-parameter draft. The published 27B figures used BF16 and eight concurrent requests.
- Limits
- Two supported target models; tested at low concurrency (eight requests); gains depend on how predictable your traffic is; the CUDA plugin tracks one SGLang version.
What’s good
- 2.71 to 3.84 times plain-decoding throughput on agent replays (self-reported)
- Beats DFlash2 on every benchmark in Tencent's tables
- Apache 2.0 code and weights; training code in SpecForge
- Output unchanged: the main model verifies every drafted token
What’s not
- Drafts for only Qwen3-8B and Qwen3.8-27B
- 27B setup needs many SGLang flags and per-workload tuning
- Hardware for the published figures not stated
- Young repository: about 234 stars, no releases
Best for
Teams serving Qwen3.8-27B or Qwen3-8B on SGLang who want more tokens per GPU without changing outputs.
DSpine alternatives
DSpine: common questions
- Is DSpine worth it?
- DSpine is a real speed-up for anyone serving Qwen3.8-27B on SGLang, and of no use yet to anyone serving anything else. We rate it 3.8 out of 5.
- How much does DSpine cost?
- Free under Apache 2.0, code and draft weights. You need your own GPU running the target model.
- What are the downsides of DSpine?
- Drafts for only Qwen3-8B and Qwen3.8-27B; 27B setup needs many SGLang flags and per-workload tuning; hardware for the published figures not stated; young repository: about 234 stars, no releases.
- Who is DSpine best for?
- Teams serving Qwen3.8-27B or Qwen3-8B on SGLang who want more tokens per GPU without changing outputs.
- How much does DSpine cost for a team or in production?
- Free software; the cost is the GPU already running the target model plus memory for a 2-billion-parameter draft. The published 27B figures used BF16 and eight concurrent requests.
- Is DSpine safe for company data?
- Runs entirely on your hardware with no telemetry described. Apache 2.0 for code and weights, with third-party components under their own licences per THIRD_PARTY_NOTICES.
- What are the best DSpine alternatives?
- The closest alternatives we have reviewed are SGLang and vLLM.
Answers checked Nov 11, 2026.