DeepSeek-V3: frontier-level open model trained for ~$5.6M in GPU time
Confirmed
Importance: historic (5 of 5)The takeaway
Chinese lab DeepSeek released DeepSeek-V3, a 671B-parameter mixture-of-experts model (37B active) with open weights that rivaled GPT-4o and Claude 3.5 Sonnet; its final training run reportedly used 2.788M H800 GPU-hours (~$5.6M).
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Historic (5 of 5)
- Last verified
- 29 September 2026
Your AI and this story
- GPT-6 AstraIn its training data
- Claude Opus 5.5In its training data
- Gemini 4 HydrogenIn its training data
- Grok 4.7In its training data
It happened before the cutoffs of all four assistants, so it can be in their training data.
Key facts
- Released 26 December 2024; technical report arXiv 2412.19437
- 671B total parameters, 37B activated per token
- Pre-trained on 14.8 trillion tokens
- 2.788M H800 GPU-hours for full training (~$5.576M at $2/GPU-hour, excluding prior research)
- Innovations: multi-head latent attention, auxiliary-loss-free load balancing, FP8 training, multi-token prediction
What happened
DeepSeek published open weights and an unusually detailed report showing frontier performance at a fraction of the reported compute of US labs, despite export controls.
Why it matters
Upended assumptions about the cost of frontier AI and China’s position; it was the base for DeepSeek-R1 weeks later.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 2 primary
Primary
- DeepSeek-V3 Technical Report (arXiv)arxiv.org, paper
- deepseek-ai/DeepSeek-V3 (code & weights)github.com, code
Changes
- Filed