The largest open-weight AI models, by parameter count
The biggest open-weight models now rival closed ones in size: some have over a trillion parameters. The table ranks every original model the site has analyzed by total parameters (GGUF re-uploads of the same models are left out).
No models yet.
Total parameters aren't the whole story
Most of the largest models are mixture-of-experts (MoE): each layer holds many small "experts", and a router picks a few of them for every token. So two numbers matter:
- Total parameters decide how much memory the model needs: every expert has to be loaded.
- Active parameters decide how much work each token takes, which is why a huge MoE model can generate text as fast as a much smaller dense one.
"On disk" is the size of the published weights. Many labs now publish in 8-bit (FP8) or 4-bit (MXFP4) formats, which is why some models with more parameters take less space than smaller ones.
Large but light
These mixture-of-experts models use at most 10 billion parameters per token: fast to run, if you have the memory to hold them.
No mixture-of-experts models that small yet.
What does it take to run them?
Each model's page has a Hardware tab with the data-center GPUs needed to hold it, and the hardware calculator shows which ones fit on a given GPU or Mac.