Generative protein language model family that jointly reasons over protein sequence, structure, and function to design novel proteins.
Currently Offered
ESM3 was introduced on 25 June 2024 as a generative language model for biology that simultaneously reasons over protein sequence, structure, and function. The largest version has 98 billion parameters, was trained on roughly 2.78 billion natural proteins using over 10^24 FLOPS of compute, and can be prompted with any combination of the three modalities to generate new proteins.
Its researchers used it to design esmGFP, a novel green fluorescent protein 58 percent similar to the closest known fluorescent protein, and published the work in Science in early 2025. The family ships in small, medium, and large sizes through a credit-metered API, with the 1.4 billion-parameter open model released under the MIT license for researchers.
