15T
median training dataset of 2025's notable AI models (29 with a published figure)
Median by year (log scale)
20122025
What the chart shows
The 29 notable models of 2025 with a reported dataset size were trained on a median of about 15 trillion datapoints. For language models a datapoint is roughly a token, a word or part of a word.
Dataset sizes span many orders of magnitude because the chart mixes text, image and game-playing models, each measured in its own units.
Key numbers
| Measure | Value |
|---|---|
| Median, 2025 (29 models) | 15T |
| Median, 2023 (63 models) | 77B |
How to read it
- Datapoints are tokens for language models, images for vision models and steps or episodes for game-playing systems, so models of different types are not directly comparable.
- Only models with a published or estimated dataset size are shown.
- Includes models published up to 30 September 2026.
Source: Epoch AI, Data on AI models (CC BY 4.0), accessed 19 October 2026. Analysis and chart by AI Adoption, following the method of the AI Index Report 2026 (Stanford HAI), figure 1.1.11. Includes models published up to 30 September 2026.
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