Grok-3

xAI · February 2025

activeClosedmixture of expertsmultimodalAPI Available
VariantsGrok-3, Grok-3 Mini

Why It Matters

Demonstrated that massive compute infrastructure (100K H100 GPUs) could produce a frontier model competitive with the best from OpenAI and Google, establishing xAI as a serious player.

Description

Trained on xAI's Colossus supercluster — one of the world's largest AI training clusters with 100,000 NVIDIA H100 GPUs. Introduced a 'Think' mode for extended reasoning (where the model works through problems step-by-step before answering) and a 'DeepSearch' mode for comprehensive research tasks. Achieved top scores on math and science benchmarks.

Notable Milestones

  • Trained on 100K H100 GPU Colossus supercluster
  • Think mode for step-by-step extended reasoning
  • DeepSearch for comprehensive research tasks
  • Top scores on math and science benchmarks (AIME, GPQA)

Benchmark Scores

GPQAGraduate-level science QA
84.6%
AIMEAMC/AIME math competition
93.3%

Key Innovations

Reasoning
ReasoningStructured step-by-step problem solving, often using chain-of-thought or tree-of-thought approaches.

Family Tree

Built On

Lineage

Grok-1Grok-2Grok-3

Successors (1)

External Links