GPT-4 (Mar 2023)
GPT-4 (Mar 2023) is an AI model developed by OpenAI (United States), first published in March 2023. It works in the multimodal, language and vision domain, on tasks such as language modeling, language modeling/generation, question answering and visual question answering. It counts among the frontier models: the systems trained with the most compute of their moment.
Training it took an estimated 2.1×10²⁵ FLOP of compute (estimation method: hardware). The model has 1,800,000,000,000 parameters. It was trained on roughly 5.4T datapoints. Training ran on 25,000 NVIDIA A100 SXM4 40 GB for about 2.3K hours. The compute alone is estimated at $37M in 2023 dollars.
Access: API access. Its weights are not openly released. The reference paper has 24,490 citations. Epoch AI rates the confidence of this record as likely.
- Organization
- OpenAI
- Country of organization
- United States
- Domain
- Multimodal, Language, Vision
- Task
- Language modeling, Language modeling/generation, Question answering, Visual question answering
- Training compute
- 2.1×10²⁵ FLOP
- Compute estimation method
- Hardware
- Parameters
- 1,800,000,000,000
- Dataset size
- 5.4T
- Training hardware
- NVIDIA A100 SXM4 40 GB
- Chips used
- 25,000
- Training time
- 2,280 h
- Chip-hours
- 57M
- Training power draw
- 19.9 MW
- Training cost (2023 USD)
- $37M
- Model accessibility
- API access
- Open weights
- No
- Citations
- 24,490
- Epoch confidence
- Likely
| 01 | Epoch Capabilities Index | ECI Score | 125.9 |