Tier 4 Deeply open
What the publisher claims: open-source, MIT licensed code and weights
The one-liner: MIT code and MIT weights over public sequence data: the practical open-weight baseline for biology.
The six questions
Code and checkpoints are downloadable and ungated on the Hugging Face Hub.
MIT licence with no field-of-use or scale restrictions; Atlas data is CC BY 4.0 for academic and commercial use.
Runs locally through PyTorch or Hugging Face transformers and is widely ported.
Corpus (UniRef50 and UniRef90), tokenizer, architecture and evaluations are published, but the full pretraining recipe and hyperparameters are not packaged.
Data and code are public so a skilled group can train a substantially equivalent model, at real scale cost.
facebookresearch/esm is archived — frozen artifacts with no development path.
License
MIT for source code and released checkpoints; ESM Metagenomic Atlas data under CC BY 4.0 with Meta Open Source Terms OSI-approved — license source ↗
- Meta Open Source Terms apply to Atlas data use
- Verify terms per checkpoint
What is actually released
- training code (fairseq based)
- pre-trained checkpoints from 8M to 15B parameters
- ESM Metagenomic Atlas structures and embeddings
- protein-design example code
Why it matters
Shows the protein-modelling openness ceiling that AlphaFold 3 weight terms fall short of, and the cleanest MIT-plus-public-data example in the batch.
Sources
- https://huggingface.co/facebook/esm2_t33_650M_UR50D
- https://huggingface.co/facebook/esm2_t36_3B_UR50D
- https://raw.githubusercontent.com/facebookresearch/esm/main/README.md
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.