Dev claims further-optimized DeepSeek V4 NVFP4 uses 190GB of 192GB VRAM
HankYeomans · x · 2026-09-22
A developer shared an experiment claiming a further-optimized DeepSeek V4 NVFP4 build that fills 190GB of 192GB total VRAM, saying he doubts he can squeeze out much more. No methods or benchmarks were provided; treat as an unverified personal claim.
More from Infra
- Meta Muse gives every user a cloud VM, and 100M users would mean a chip problem — AccBalanced · 2026-09-22
- DeepSeek bets next model on Huawei chips after Ascend 910C training run failed — soumitrashukla9 · 2026-09-22
- Inference era shifts to 1-30MW micro datacenters as OpenAI and Anthropic chase distribution — AccBalanced · 2026-09-22
- World model startup Reactor raises $59M to build the developer platform for world models — buckymoore · 2026-09-22
- Meta unveils Petal, first petabit-class transoceanic subsea cable, live in 2029 — Meta_Engineers · 2026-09-22
- One of the biggest AI launches ever stayed up under crazy load, out-uptimeing Anthropic — hardimanjames · 2026-09-22