AIster · The model wars · Launches
Chinese lab DeepSeek releases DeepSeek-V3
DeepSeek says it rivals closed models, with official training priced at $5.576 million
Key facts
- Date
- 26 Dec 2024
- Era
- The model wars
- Interest
- Launches
- Threads
- Open vs closed, Compute
- Who
- DeepSeek
- Size
- 671B parameters, 37B active per token; 14.8T training tokens
- Training cost
- 2.788M H800 GPU hours, $5.576M at an assumed $2/GPU-hour; excludes prior research and ablations
The Chinese lab DeepSeek released V3, a 671-billion-parameter open-weights model (its model files are published for download) it said performed comparably to leading closed models. Its paper put the official training at 2.788 million H800 GPU hours (time on NVIDIA's H800 chips), or $5.576 million at an assumed rental price of $2 an hour, excluding earlier research and experiments. A month later its reasoning model R1 helped wipe nearly $600 billion off NVIDIA's market value in a day.
Sources
- arXiv (DeepSeek-AI), DeepSeek-V3 Technical Report, 2024-12-27
- DeepSeek (archived), Introducing DeepSeek-V3, 2024-12-26
- TechCrunch, Nvidia drops $600B off its market cap amid the rise of DeepSeek, 2025-01-27