Grok-3

model
SpaceXAI

Trained on 200K GPUs at the Colossus supercomputer with 10x more compute than Grok-2 and 12.8T training tokens. MoE architecture, estimated ~3T parameters. Up to 1M token context via API. Features Think mode (extended reasoning) and DeepSearch (agentic web research).

Grok-3 mini Reasoning scores AA Intelligence Index v4.3: 15. AA Intelligence Index v4.3: 12. Proprietary.

AnnouncementArtificial AnalysisOpenRouter

Model Details

Architecture MOE
Parameters (est.) ~ 2.1T
Context window 1,000,000
AA Intelligence 12 was 12 on v4.2

Variants

Name Parameters Notes
Grok-3 — —
Grok-3 mini — Reasoning variant, AA Intelligence Index v4.3: 15
frontierreasoningmoe

Related