Grok 3
AI Model
Language
Prior generation
What 200,000 H100s buys: Colossus's first frontier model.
You're looking at an AI model record: who built it, what it can do, and the hardware and data behind it.
Grok 3 (February 2025) was trained on the Colossus cluster with roughly ten times Grok 2's compute (Epoch-class estimates put it near 4–5×10²⁶ FLOP) and arrived with a Thinking mode competitive with o-series reasoning — xAI's proof that raw build speed converts to frontier capability.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Developed by
Trained on
Succeeded by
Sources
Record reviewed as of 2026-01-15 — see Methodology.