Qwen3.8-27B-DSpine-b16: DSpine Draft Model Boosts Qwen3.8-27B Inference Speed by 3.8x

A draft model for speculative decoding dedicated to Qwen3.8-27B. It consists of 5 layers, with each layer passing predicted prefix features from adjacent positions to propagate causal dependencies through network depth. It supports block size 16 and an 8,192 token sliding window.

Evaluated in the SGLang 0.5.22 environment with BF16 precision. Compared to the existing DFlash2, it increases acceptance length by up to 1.5 tokens on general benchmarks and achieves over 20% improved acceptance length on agent tasks. Notably, it achieved an acceptance length of 5.63 tokens on the WorkBuddy Office benchmark during xhigh inference.

In terms of end-to-end throughput, it shows up to a 3.84x speed improvement over autoregressive decoding. It recorded 805.53 tokens/s in WorkBuddy Office low inference mode, and acceleration effects of over 2.9x were also confirmed on SWE-bench Pro and τ²-bench. For dynamic verification, it references DSpark and enables ReplaySSM to ensure stability.

Distributed under the Apache 2.0 license and composed of BF16 tensors of 2B parameter size. It can be applied immediately by specifying speculative-algorithm as DSPINE and registering the corresponding checkpoint path in the SGLang server configuration. It is suitable for developers looking to reduce inference latency for Qwen3.8-27B-based services.