In brief: These models solve overlapping but different parts of protein science. Start with the task: design, representation or structure prediction.
At a glance
| Model | Primary role | Typical input | What to validate |
|---|---|---|---|
| ESM3 | Conditional, multimodal generation | Partial sequence / structure / function information | Viability, folding and actual function |
| ESMC | Protein representations and embeddings | Sequence | Downstream-task accuracy and leakage |
| ESMFold2 | High-resolution structure prediction | Sequence; optional MSA where supported | Geometry, interfaces and biological relevance |
For researchers designing new proteins
ESM3 is the natural place to explore prompted generation. A design still needs selection, structural assessment and empirical testing. Verify which ESM3 variant is actually available to your account; not all checkpoints and hosted endpoints are equivalent.
For embedding and classification pipelines
ESMC provides sequence representations suitable for analysis and downstream models. Building a classifier on embeddings can be simpler and more reproducible than attempting to fine-tune a large generative model. Split homologous families carefully to avoid overly optimistic performance.
For structure prediction
ESMFold2 is designed for folding and complex structure hypotheses. A predicted structure does not capture all protein dynamics or guarantee that a proposed binding interaction occurs in vitro or in vivo.
What about licensing and costs?
Review each current model card, checkpoint license and hosted API pricing independently. Open model weights do not necessarily mean every hosted service or downstream dataset is available under identical terms.
Primary sources and further reading
Editorial review: October 2026. This guide is educational. Model availability and APIs may change; confirm exact versions in the linked documentation. Generated structures and sequences require independent scientific validation.