Who still tells you how big their model is
In 2021 and 2022, 10 of the 11 frontier models whose size could be traced to the people who built them were models you could not download. Between 2023 and 2025, that number is 0.
Closed laboratories used to tell you how large their models were. Gopher was 280 billion parameters, Megatron-Turing 530 billion, PaLM 540 billion, and all three were announced with the figure attached by the companies that trained them. None of those weights were ever released. Publishing the number and keeping the model were, for a few years, entirely compatible.
That stopped. Of 1,072 notable models in this dataset, 725 carry a parameter count of some kind, but only 417 carry one that the people who maintain the dataset were willing to trace to a developer, a paper or a model card. The gap between those two numbers is where this exhibit lives.
| Year | Frontier | Size traceable | Open | Closed |
|---|---|---|---|---|
| 2017 | 3 | 2 | 0 | 1 |
| 2018 | 1 | 1 | 1 | 0 |
| 2019 | 10 | 9 | 5 | 4 |
| 2020 | 4 | 4 | 1 | 3 |
| 2021 | 8 | 7 | 1 | 6 |
| 2022 | 5 | 4 | 0 | 4 |
| 2023 | 9 | 1 | 1 | 0 |
| 2024 | 11 | 2 | 2 | 0 |
| 2025 | 6 | 0 | 0 | 0 |
Frontier means Epoch's own Frontier model flag. The flag is applied in review, so the most recent year is usually incomplete and a zero there means not yet classified rather than none built.
The figures did not stop circulating when they stopped being published. These 7 closed frontier models all have a parameter count attached in the dataset, and not one of those counts is traceable to anybody who has seen the model. The note beside each figure is quoted exactly as the dataset records it.
- Grok 33.0TxAI · 2025 · recorded as likely
No note recorded at all.
- Grok 43.0TxAI · 2025 · recorded as speculative
“Rumored to be 2.4T params (https://x.com/kalomaze/status/1942996555088134592)”
- Llama 4 Behemoth (preview)2.0TMeta AI · 2025 · recorded as likely
“Llama 4 Behemoth, a 288 billion active parameter model with 16 experts that is our most powerful yet and among the world’s smartest LLMs.”
- Amazon Titan200BAmazon · 2023 · recorded as likely
“200B dense model https://importai.substack.com/p/import-ai-365-wmd-benchmark-amazon”
- GPT-4 (Jun 2023)1.8TOpenAI · 2023 · recorded as likely
“Rumored to be 1.8T parameter MoE with 280B activated on the forward pass, per https://www.semianalysis.com/p/gpt-4-architecture-infrastructure. Other sources estimate 1.76T with 220B per forward pass https://web.archive.org/web/20230712123915/https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/”
- GPT-4 (Mar 2023)1.8TOpenAI · 2023 · recorded as likely
“Rumored to be 1.8T parameter MoE with 280B activated on the forward pass, per https://www.semianalysis.com/p/gpt-4-architecture-infrastructure. Other sources estimate 1.76T with 220B per forward pass https://web.archive.org/web/20230712123915/https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/”
- PaLM 2340BGoogle · 2023 · recorded as likely
“Model Architecture: "PaLM-2 is a new state-of-the-art language model. We have small, medium, and large variants that use stacked layers based on the Transformer architecture, with varying parameters depending on model size. Further details of model size and architecture are withheld from external publication." However, the parameter count was leaked to CNBC: https://www.cnbc.com/2023/05/16/googles-palm-2-uses-nearly-five-times-more-text-data-than-predecessor.html”
Two of these deserve naming. The figure most often repeated for GPT-4 is sourced to a newsletter rather than to OpenAI. The figure for Grok 3 has no stated source of any kind, and the one for Grok 4 is recorded as 3T while the note beneath it quotes a rumour of 2.4T.
A model counts as sized when Epoch records a parameter count and marks its confidence Confident, which is their assessment that the figure is traceable to the developer, a paper or a model card.
Figures marked Likely or Speculative are counted separately and never as disclosure. Those are reconstructions, and several are sourced to a single social media post.
What this does not show
An absent confident figure is not proof a developer said nothing. It is proof that Epoch, who maintain this dataset, could not trace a figure to a source they were willing to stand behind, and that judgement has false negatives: Kimi K3 is marked speculative at 2.8T while Moonshot's own model card states the figure.
It is also a count of models rather than of laboratories, and “notable” is the dataset’s own inclusion criterion rather than ours. A lab that published one figure and withheld ten appears here once on each side.
Every figure on this page comes from one dataset, unmodified except by the counting described above. It is published under CC BY 4.0, which is what makes reproducing it here permitted, and requires the credit below.
Epoch AI, 'Data on AI Models'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/data/ai-models-documentation'
https://epoch.ai/data/ai-models-documentation
Re-measured from the source on a schedule, most recently on 23 September 2026. This reading was taken after the current build was deployed. The counting is done by scripts/museum-data.mjs, which refuses to publish a reading it cannot validate, so a figure here may be older than today but is never a guess.