antirez streams a half-trillion-param DeepSeek v4.1 Flash on one DGX Spark at ~9 t/s

antirez · x · 2026-09-14

antirez demonstrated running DeepSeek v4.1 Flash (half a trillion parameters) on a single DGX Spark via SSD streaming: 9 t/s generation, or 22 t/s with dual-Spark RDMA. He calls it "kinda usable for QA" but not for coding unless left running overnight. The support landed in his DwarfStar project.

Related event: antirez Runs 500B-Param DeepSeek on a Single DGX Spark via SSD Streaming(3 posts)→

Original post →

More from Infra

Infra channel →