deepseek-v4.1POST

vLLM and SGLang match or beat DeepSeek's open-sourced inference engine on V4.1 Flash

Four days after DeepSeek open-sourced its inference engine, vLLM claims parity with DeepInfer on V4.1-Flash decode and SGLang claims up to 7.4% faster with a repro guide.

Aaron

@dicemanoramaArchived 6 days agoOriginally posted 2026.10.25

Aaron @dicemanoramaOriginally posted 2026.10.25

Open original on X

DeepSeek open-sourced its in-house inference engine four days ago. vLLM and SGLang already caught it. vLLM says it matches DeepInfer on decode for V4.1-Flash. SGLang says up to 7.4% faster, with a repro guide. Share the good stuff and the whole field speeds up. 🚀 https://x.com/sgl_project/status/2114071115605877043

This post quotes another post: x.com/sgl_project/status/2114071115605877043

x.com/dicemanorama/status/2114110723567874380Archived 2026.10.25 12:02