Tier 3 Practical open-weight
What the publisher claims: open, MIT
The one-liner: A genuinely open small checkpoint sitting beside a closed flagship.
The six questions
The 1.4B open checkpoint is downloadable and ungated as of this snapshot.
MIT terms now apply to the open weights and the code.
Runs locally through the esm package or through transformers.
The paper documents the corpus (2.78B natural proteins plus synthetic augmentation), tokenization and filtering, but the training dataset is not published.
The open checkpoint can be retrained at small scale from published method detail; the flagship cannot.
The Biohub/esm SDK accepts outside PRs and publishes tutorials, though direction is institutional.
License
MIT for code and the open checkpoint (Biohub/esm LICENSE.md and ESM3 README); the 2024 release shipped a non-commercial Community License requiring acceptance on the Hub OSI-approved — license source ↗
- Earlier snapshots were gated behind non-commercial terms; check the exact release you cite
What is actually released
- ESM3 code and SDK
- esm3-sm-open-v1 weights
- cookbook tutorials
- model card with safety filtering notes
Why it matters
Illustrates the common pattern where openness is granted for a reduced model while the capable model stays behind an API.
Sources
- https://huggingface.co/EvolutionaryScale/esm3-sm-open-v1
- https://raw.githubusercontent.com/Biohub/esm/main/LICENSE.md
- https://raw.githubusercontent.com/Biohub/esm/main/_assets/ESM3_README.md
- https://www.evolutionaryscale.ai/blog/esm3-release
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.