ESM3 / EvolutionaryScale vs ESM-2 / ESMFold

ESM3 / EvolutionaryScale: Multimodal generative protein language model over sequence, structure and function (ESM3; also ESM C embeddings). ESM-2 / ESMFold: Protein language models and single-sequence structure prediction (Meta repo archived; models still widely used via Hugging Face).

Facts below come from each project's own website, repository and paper. We don't publish benchmark claims here; see the papers for evaluations. Shared tags: plm.

FieldESM3 / EvolutionaryScaleESM-2 / ESMFold
TaskGenerative design, embeddings, structure predictionProtein language model, structure prediction
ModalitiesProtein sequence, structure and function tracksProtein (single chain)
MSA requiredNo (single-sequence generative masked LM)No (single sequence)
Open weightsESM3-open: yes (MIT); larger models via APIYes
GPU needCUDA or CPU (esm3-sm-open-v1)—
LicenseMITMIT
PricingNo-cost tier + paid plansOpen source
Released2024-07-022022-07-21
PaperLinkLink
GitHubGitHubGitHub
GitHub stars2,9934,170
InterfacePackagePackage, Colab

License and commercial use

FieldESM3 / EvolutionaryScaleESM-2 / ESMFold
Code licenseMITMIT
Weights licenseMIT—
Commercial usePermissive (commercial OK)Permissive (commercial OK)
AnswerYes. The recorded license is permissive, so closed-source commercial use is allowed if you keep the notices. DetailsYes. The recorded license is permissive, so closed-source commercial use is allowed if you keep the notices. Details

Add more Protein structure & design tools to this comparison →

More Protein structure & design comparisons

← All Protein structure & design tools