Protein language model family for representation learning, providing embeddings and zero-shot mutation analysis for proteins.
Currently Offered
ESMC, announced in December 2024 as ESM Cambrian, is the representation-learning family of the ESM model line, built in parallel to the generative ESM3 flagship. It serves as a drop-in replacement for earlier ESM2 models with claimed linear scaling, with the 300M model matching ESM2 650M and the 6B version setting new representation benchmarks.
The current 2026 release is served through the Biohub Platform API or run locally from open weights released under the MIT license in 300M, 600M, and 6B sizes. Sparse autoencoders trained on ESMC decompose its representations into roughly 16,000 interpretable features used in the ESM Atlas, supporting structure prediction, function annotation, protein design, and evolutionary analysis.
