Grok 3

Training compute

3.5×10²⁶ FLOP

Parameters

3T

Published

Feb 17, 2025

Grok 3 is an AI model developed by xAI (United States), first published in February 2025. It works in the language, vision and multimodal domain, on tasks such as chat, language modeling/generation, question answering and 2 more. It counts among the frontier models: the systems trained with the most compute of their moment.

Training it took an estimated 3.5×10²⁶ FLOP of compute (estimation method: hardware,comparison with other models). The model has 3,000,000,000,000 parameters. Training ran on 80,000 NVIDIA H100 SXM5 80GB for about 2.2K hours. The compute alone is estimated at $218M in 2023 dollars.

Access: API access. Its weights are not openly released. Epoch AI rates the confidence of this record as likely.

Full record

Organization: xAI
Country of organization: United States
Domain: Language, Vision, Multimodal
Task: Chat, Language modeling/generation, Question answering, Code generation, Visual question answering
Training compute: 3.5×10²⁶ FLOP
Compute estimation method: Hardware, Comparison with other models
Parameters: 3,000,000,000,000
Training hardware: NVIDIA H100 SXM5 80GB
Chips used: 80,000
Training time: 2,160 h
Training power draw: 109.9 MW
Training cost (2023 USD): $218M
Model accessibility: API access
Open weights: No
Epoch confidence: Likely

Benchmark results

01	Epoch Capabilities Index	ECI Score	139
02	WeirdML	Accuracy	0.37
03	SimpleBench	Score (AVG@5)	0.36
04	ARC-AGI	Score	0.06
05	Apex Agents	Pass@1 score	0.02
06	ARC-AGI-2	Score	0