Fork builds engine to run Qwen at ~400 tok/s on three used V100s
jfinkpottery · reddit · 2026-10-27
Reddit user jfinkpottery built sfnx, an open-source inference engine that runs Qwen-Flash-Next at nearly 400 tokens/sec on three used V100 32GB cards inside an old Dell R740. Forked from strata but with engine internals rewritten for the old Volta generation, it shows a cheap path to serving new models on secondhand hardware.
More from Infra
- Deep dive: how Weka is breaking Nvidia's AI memory bottleneck — AccBalanced · 2026-10-27
- Anthropic locked into $475B compute deals, OpenAI lifts projected spend to ~$750B through 2030 — kyliebytes · 2026-10-27
- First systematic data center noise study: 17 of 25 hyperscalers drew no credible complaints — AndyMasley · 2026-10-27
- RFD4-Proteina runs on an RTX 3060: protein design from $0.001 per design — AllThingsApx · 2026-10-27
- NPM packages now take ~20 minutes to go live, dev suspects AI-attack throttling — cocktailpeanut · 2026-10-27
- Nature Physics: reprogrammable refractive index enables arbitrary multimode wave control for ML — MarioKrenn6240 · 2026-10-27