Rumor: Grok 5 Targets 6T Params, 10T Model in the Works
Hesamation · x · 2026-08-20
A leak suggests Grok 5 is in training targeting 6 trillion parameters (Grok 4 is 3T, Kimi K3 is 2.8T). Training occurs on the Colossus II cluster with 220k high-end GPUs and 1GW power draw. Elon confirmed it will use 25 years of SpaceX engineering data and arrive before late 2026. The source also hints at a separate 10T parameter model being developed.
More from Infra
- AirLLM update runs 2.8T Kimi K3 on under 4GB VRAM — pmttyji · 2026-08-20
- Cloudflare fixes remote Spectre attack in Workers — ifsecure · 2026-08-20
- Why LM Studio is slower than llama.cpp: You aren't offloading ALL layers to GPU — Fancy-Snow7 · 2026-08-20
- LLM Latency Optimization: Why Tripling GPU Power Barely Improves TTFT — HankYeomans · 2026-08-20
- Intel ships OpenVINO 2026.3 with more models, new GenAI pipelines and memory optimizations — emmanuelvivier · 2026-08-20
- each::labs releases Video API with 39 models for post-processing and packaging — thetripathi58 · 2026-08-20