ESM-3

ESM-3 is EvolutionaryScale's frontier generative model for biology, released in 2024. It jointly reasons across the three fundamental properties of proteins, sequence, structure, and function, represented as discrete token tracks that can be partially prompted and iteratively unmasked to design new proteins.

At its largest scale, ESM-3 trained with over 1e24 FLOPs on 2.78 billion proteins and 771 billion unique tokens, reaching 98 billion parameters, and the work was published in Science. Only the 1.4B esm3-sm-open-v1 variant is openly downloadable (under a non-commercial license); the 7B and 98B models run behind EvolutionaryScale's Forge API.