antirez: 1 token/sec local Kimi K3.1 inference needs 128GB memory, 512GB much better
antirez · x · 2026-11-08
antirez adds the memory requirements for local Kimi K3.1 inference: 1 token/sec generation needs at least 128GB of memory, while 512GB makes things a lot better.
Related event: antirez's DwarfStar runs Kimi K3.1 PRO locally from safetensors(4 posts)→
More from Infra
- Colossus II nearly triples compute to 1.7 GW in a year, targeting 1.2M Blackwell GPUs — XFreeze · 2026-11-08
- Polymarket odds: 27% chance a state enacts data center moratorium this year — Polymarket · 2026-11-08
- Polymarket: Anthropic reportedly agrees to $3.4B deal for gas-fired generators — Polymarket · 2026-11-08
- NVIDIA open-weights Nemotron 3.5 Transcribe: 30B-param speech model in 40 languages — dl_weekly · 2026-11-08
- Musk mocks $150k Mac Studio local LLM setup that's worse than a $200/mo subscription — elonmusk · 2026-11-08
- Bank dev tests 1.7B local model offline: all traffic stayed on localhost, emails in 2.5s — Ok_Resolve4741 · 2026-11-08