本地部署 V4.1 Flash：2.9 bit 量化，512K 上下文 126 tok/s

Neko Legends @softpoo原推发布于 2026.10.21

在 X 上打开原推

Local now serving DeepSeek-V4.1-Flash at 2.9 bits with 512K of context per request, on our own inference engine. One stream: 126 tok/s writing code. Eight streams: 296 tok/s total. Reading a 21K+-token prompt: 5,000 tok/s Rebased from Tensorfold 1.1.0, will make a PR back in to @ashxhart official repo after a few more extra fixes and improvements due to newly released DeepInfer (not entirely compatible but 4 concepts can be borrowed). .