SGLang adds DFlash to push Qwen 3.5 397B-A17B inference up to 4.3x faster

Z Lab, Modal and LMSYS released a DFlash drafter for Qwen's 397B model and benchmarked it above native MTP on 8x B200 GPUs.

By · Published

Primary source: LMSYS Org

Why it matters

The release shifts the open-model fight from model access to serving economics: faster verified decoding can lower the cost of running frontier-scale open weights.

SGLang adds DFlash to push Qwen 3.5 397B-A17B inference up to 4.3x faster — Z Lab, Modal and LMSYS released a DFlash drafter for Qwen's 397B model and benchmarked it above native MTP on 8x B200 GPUs.

Jian Chen (@jianchen1799), Yesheng Liang and Zhijian Liu (@zhijianliu_) have pushed DFlash, Z Lab's block-diffusion speculative decoding method, into SGLang through a collaboration with Modal and LMSYS, with the teams reporting up to 4.31x higher throughput over a non-speculative baseline on Qwen 3.5 397B-A17B.…

Reader comments

Conversation for this story loads after sign-in.