SGLang's DeepSeek V4.1 Flash speed claims on GB300 and B300 GPUs, all self-reported
Article timeline
The articles in this story. Times are ET.
SGLangOfficialSGLang explains how it matched DeepInfer on prefill and decodeAISGLang's account of matching DeepInfer's throughput is out as a two-minute explainer covering the benchmarks and the DeepSeek kernels now in SGLang. Its earlier post reports up to 1.04× prefill throughput on 4× GB300 and 4× B300, and up to 7.4% decode throughput on 64× GB300. The post also links to an end-to-end reproduction guide.
- SGLangOfficialSGLang doubles DeepSeek V4.1 Flash low-latency and throughput speeds
AISGLang says its MegaKernel optimizations raise DeepSeek V4.1 Flash interactivity from 538 to 943 tokens per second per user. Its WideEP and system-wide optimizations raise total throughput from 348K to 654K tokens per second per GPU, counting cache, new input, and output tokens. The project says further improvements are still possible and promises more updates soon.
LMSYS OrgOfficialLMSYS says SGLang matches DeepInfer on DeepSeek-V4.1-Flash prefill and decode throughputAILMSYS Org says SGLang matches DeepInfer on both prefill and decode for DeepSeek-V4.1-Flash. It reports up to 1.04× prefill throughput on 4× GB300 and 4× B300, and up to 7.4% decode throughput on 64× GB300. The post also links an end-to-end reproduction guide and a production serving stack.
- SGLangOfficialSGLang matches DeepInfer on prefill and decode throughput
AISGLang says it now matches DeepInfer performance on both prefill and decode. It reports up to 1.04× prefill throughput on 4× GB300 and 4× B300, and up to 7.4% decode throughput on 64× GB300. The project also provides an end-to-end reproduction guide and a full serving stack for production use.
