Qwen3.8-27B-DFlash2: DFlash 2 Boosts Qwen3.8-27B Inference Speed by 3.4x

DFlash 2 is a draft model designed to accelerate the inference speed of the Qwen3.8-27B model. It does not operate independently but serves as an auxiliary component for speculative decoding within inference servers such as SGLang or vLLM. It pre-generates tokens for the target model to verify.

As a block diffusion-based drafter, it predicts entire token blocks in a single pass. By applying two-tab dynamic synthesis to the backbone, it prevents the degradation of draft quality towards the latter part of the block. This is a lossless decoding method where greedy outputs match the target model exactly, and the distribution is preserved during sampling.

Benchmarks conducted on a single NVIDIA H200 GPU using SGLang show a throughput improvement of up to 3.43x compared to autoregressive decoding. These figures are higher than those achieved by Qwen3.8's built-in MTP or the community DSpark drafter. Consistent performance improvements were observed across various benchmarks, including GSM8K, MATH-500, and HumanEval.

Released under the Apache-2.0 license, it is a BF16 model with 2B parameters. It can be applied in environments serving Qwen3.8-27B to reduce response latency. It can be used immediately by simply specifying the draft model path in existing inference server configurations.