| name | esmc |
| description | ESM C / ESM Cambrian (ESMC) — next-generation protein representation language models scaling beyond ESM-2 (300M / 600M open weights; larger tiers may need Forge/API). Use when: (1) Stronger AA embeddings than ESM-2 for metagenomic proteins, (2) Representation learning (not ESM3-style controllable generation), (3) Downstream heads that benefit from Cambrian checkpoints. Upstream: https://github.com/evolutionaryscale/esm (Biohub/esm). Route via protein-language-model. Keep ESM-2 workflows on esm when pinned for reproducibility.
|
| license | MIT |
| category | analysis-tools |
| tags | ["ESMC","ESM-C","ESM-Cambrian","PLM","protein-language-model","EvolutionaryScale"] |
| upstream | https://github.com/evolutionaryscale/esm |
| stage | function |
ESM C (ESM Cambrian / ESMC)
Upstream: evolutionaryscale/esm ·
Blog: https://www.evolutionaryscale.ai/blog/esm-cambrian ·
Biohub: https://biohub.org/blog/esm-cambrian-unsupervised-learning
Citation
Candido, S. et al. Language Modeling Materializes a World Model of Protein
Biology. bioRxiv https://doi.org/10.64898/2026.06.03.729735 (2026).
Interim blog (if preprint not yet cited in Methods): ESM Team. ESM Cambrian:
Revealing the mysteries of proteins with unsupervised learning (2024).
https://www.evolutionaryscale.ai/blog/esm-cambrian
See also docs/references.md.
Analytical thinking
ESM C is the representation-learning sibling of generative ESM3. It
targets embeddings / biological understanding with a new scaling frontier vs
ESM-2. Open weights commonly discussed: 300M and 600M; larger
(e.g. 6B) tiers may require Forge / commercial endpoints — pin license and
access path.
| vs | Prefer ESMC when |
|---|
esm (ESM-2) | New embedding jobs that can adopt Cambrian checkpoints |
saprot | Sequence-only representation; not Foldseek SA-token PLM |
prostt5 | Embeddings / property heads — not AA↔3Di for Foldseek DB build |
| ESM3 | Representation vs multimodal controllable generation |
Do not silently swap ESM-2 for ESMC in published PLMSearch/LucaPhylo-style
pipelines without re-benchmarking. Check academic vs commercial terms for each
size.
How to run
Do not invent API method names — follow the installed esm package docs.
Decision tree
Need modern protein embeddings?
├─ ESM C / Cambrian checkpoints → esmc
├─ Legacy / pinned ESM-2 pipelines → esm
├─ Structure-aware AA+3Di tokens → saprot
├─ Foldseek 3Di from FASTA → prostt5
└─ Remote homology wrappers → plmsearch / deepblast
Related skills
protein-language-model · esm · saprot · prostt5 · plmsearch ·
deepblast · foldseek · metagenomics-llm · tool-selection
