Skip to content
← Models
Open weightsEvolutionaryScale/10 trustMetrics-checked

ESM3-open

EvolutionaryScale's open generative protein model: designs proteins by jointly reasoning over sequence, structure and function.

Visit official page →

About

ESM3 is the flagship of EvolutionaryScale, the biology AI lab founded by the former Meta ESM team, launched in June 2024. It departs from the encoder-only ESM-1/ESM-2 line: a masked generative transformer that represents a protein across three simultaneous tracks — amino-acid sequence, discrete structure tokens, and function annotation — so a single model can be prompted with any combination of partial sequence, structure and function and asked to complete the rest. Trained on 2.78 billion proteins, it famously generated a novel green fluorescent protein that the team estimates would take 500 million years of natural evolution to arise.

The open release, esm3_sm_open_v1 (1.4B parameters), is free for non-commercial scientific use under the Cambrian Open License; the 7B and 98B flagships are served through the Forge API. The underlying science is published in Hayes et al., Science 2025. For academic protein-design work it is the most capable openly downloadable generative biology model, though commercial teams need the paid Forge tier.

Specs

Context window

Not applicable — protein sequence/structure/function input, not token context

Pricing

Free / open weights for non-commercial scientific use; paid Forge API for larger models

Free quota

Free / open weights — self-host at no license cost (non-commercial)

API access

Hugging Face (EvolutionaryScale/esm3, esm3_sm_open_v1); Forge API for 7B/98B models; pip-installable esm package

Weights

Open weights

Released

Not disclosed

Benchmarks

  • No standard benchmark table; Science 2025 paper (Hayes et al.) reports generative design results including novel GFP generation (peer-reviewed, provider-run)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

protein-designdrug-discoverybioinformaticsresearch

Trust score — 8/10

/10 trustMetrics-checked

Scored 8/10 (metrics tier, metrics-v1). Evidence: EvolutionaryScale launch Jun 2024 with peer-reviewed Science 2025 paper (Hayes et al.); 1.4B open checkpoint under Cambrian Open License; trained on 2.78B proteins; NVIDIA BioNeMo/NIM distribution. Open weights. Facts corroborated across Business Wire, NVIDIA's blog, and community tooling docs. Strengths: peer-reviewed generative biology claims, genuinely novel three-track architecture, real open checkpoint. Docked for: (1) Cambrian Open License is non-commercial — not OSI-open; (2) the open 1.4B is far smaller than the 98B flagship, so published SOTA results don't fully transfer to the downloadable weights.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
8.0/10
Openness · license + docs
6.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · EvolutionaryScale launch 2024-06-25; peer-reviewed Science 2025 paper; 1.4B open checkpoint; trained on 2.78B proteins
  • · Docked: Non-commercial license — not OSI-open
  • · Docked: Open 1.4B far smaller than 98B flagship; SOTA results don't fully transfer

Sources