deepseek-v4.1POST
DeepSeek V4.1 Flash 全本地部署：双 DGX Spark 达 50-66 tok/s
用 2 台 DGX Spark 加一台 x86 机器全本地跑 V4.1 Flash，单请求 50-66 tok/s、合计约 105-119 tok/s，prefill 2200 tok/s，KV 2.5M。
Nick Mermiges, Esq. 🦁🏈
@NicoTheGreco7 天前收录原推发布于 2026.10.24
Nick Mermiges, Esq. 🦁🏈 @NicoTheGreco原推发布于 2026.10.24
DeepSeek V4.1 Flash, fully local: 2 DGX Sparks + an x86 box. Drafter on 2 NVLinked 3090s, Engram in its RAM via RDMA. ~50-66 tok/s per request, ~105-119 combined, 2,200 tok/s prefill, 2.5M-token KV. https://github.com/ProprietaryLegal/deepseek-v41-flash-2x-dgx-spark-progress
x.com/nicothegreco/status/2113774487220949433采集于 2026.10.24 12:07