Tencent Open-Sources DSpine, a Parallel Draft Model for Qwen3.8-27B: Vendor Figures Show About 3.19x Throughput on SWE-bench Pro Replay, Ahead of DFlash2 Across the Board

Core Background & Industry Pain Points

LLMs decode one token at a time and run the whole network for every token. For coding agents, with long outputs and many tool-call turns, that directly sets latency and how many requests each GPU can serve. Speculative decoding has a small draft model guess several tokens at once, then lets the large model verify them in one pass and keep the correct ones, without changing the output distribution. At 14:20 UTC on November 7, Tencent published tencent/Qwen3.8-27B-DSpine-b16 on Hugging Face, a draft model built specifically for Alibaba's open Qwen3.8-27B. Local-inference user Lasse (@lasse_ai) shared it on X saying it beats DFlash2 in every category; the post had about 148 likes and 120 bookmarks at capture.

Architecture Highlights & Underlying Mechanisms

According to the model card, DSpine generates draft tokens in parallel: each layer passes predicted predecessor features to adjacent positions, carrying causal dependencies through network depth instead of stepping forward one token at a time like an autoregressive drafter. This checkpoint is a five-layer draft model with about 2.03B parameters, block size 16 and an 8,192-token sliding window. Inference uses dynamic verification that references DSpark, with ReplaySSM enabled. Code is on GitHub at Tencent/DSpine, and the paper, "Draft in Parallel, Condition Through Depth", is arXiv 2609.36173. Tencent had already released a Qwen3-8B version in October.

Benchmarks & Measured Results

All figures below are vendor figures, measured with SGLang 0.5.22, BF16, FlashAttention 3 and concurrency 8. On a SWE-bench Pro replay at xhigh reasoning, autoregressive decoding ran at about 248 tokens/s, DFlash2 at about 712 (2.87x) and DSpine at about 791 (3.19x); on WorkBuddy Office at low reasoning DSpine reached 3.84x. Acceptance length was 4.77 vs 4.00 on SWE-bench Pro at xhigh and 7.20 vs 5.69 on GSM8K with greedy decoding. Note that DSpine drafts 16 tokens per block while the official DFlash2 drafts 8, so the setups are not equivalent; each agent subset replays only 20 conversations, and there is no third-party reproduction yet.

Developer Hands-on Guide

Teams already running Qwen3.8-27B coding agents locally or on their own clusters can try it directly: clone Tencent/DSpine and run pip install -e ., then start sglang serve with --speculative-algorithm DSPINE and --speculative-num-draft-tokens 16 pointing at the draft model path; the model card lists the full flags, including ReplaySSM and Mamba cache options. The draft model adds about 4GB of VRAM. Speculative decoding does not change outputs, but the real speedup depends on your prompt mix and concurrency, so replay your own agent logs and compare before replacing DFlash2 or EAGLE-style drafters.