Esm

This gives you EvolutionaryScale's protein language models for generative design and embeddings. ESM3 handles the generative side: you can mask parts of a sequence and have it fill in the gaps, do inverse folding to design sequences for target structures, or condition on function annotations. ESM C is for embeddings when you need representations for downstream ML tasks. The small ESM3 model runs locally, but the better ones (7B and 98B parameters) require their Forge API. If you're doing any serious protein engineering or need to extract features from sequences, this is the current state of the art. The chain-of-thought generation for iterative refinement across sequence, structure, and function is genuinely useful for design work.