DeepSeek 员工回应第三方用 vLLM 部署 V4.1 Flash 并复现结果

Yifan Qiao @yifandotqiao原推发布于 2026.10.25

在 X 上打开原推

Nice writeup! So glad to see a third party serving DeepSeek-V4.1-Flash with vLLM and reproducing our results! Also love how the blog identifies DeepInfer's Engram RDMA fetch bottleneck, which we've observed and taken seriously as well. Regardless, vLLM is the full-fledged serving engine that takes care of the heavy lifting. Talk is cheap. Try it yourself!

这条是引用着另一条写的：x.com/kbesplv/status/2112913887817355463