597 t
estimated training emissions of DeepSeek v3 (671B parameters), against 8,930 t for Llama 3.1 405B (405B)
What the chart shows
Larger models generally produce more emissions in training, but size is not the whole story. DeepSeek v3, with 671 billion parameters, produced an estimated 597 tonnes of CO₂ equivalent, about a fifteenth of the 8,930 tonnes for Llama 3.1 405B, with 405 billion.
Hardware efficiency, training duration and the carbon intensity of the electricity used all shift a model’s footprint.
Key numbers
| Measure | Value |
|---|---|
| DeepSeek v3: 671B parameters | 597 t |
| Llama 3.1 405B: 405B parameters | 8,930 t |
| GPT-4: 1.8T parameters (est.) | 5,184 t |
| Grok 4: 3T parameters (est.) | 72,816 t |
How to read it
- Estimates by the AI Index for 13 select models.
- Both axes are logarithmic.
- Parameter counts for some closed models (GPT-4, Grok 3, Grok 4) are third-party estimates.
Source: AI Index, 2026, via the AI Index Report 2026 (Stanford HAI), figure 1.4.4: Estimated carbon emissions and number of parameters by select AI models. Chart and text by AI Adoption.
Numbers checked against the published data on 19 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 19 October 2026.
Download the data (CSV) · Cite this chart
Discussion