Skip to content
← Models
Open weightsEvolutionaryScale (now Biohub)/10 trustVerified
ESM3 logo

ESM3

A generative protein language model that reasons over sequence, structure, and function.

Visit official page →

About

ESM3 is a 98-billion-parameter multimodal protein model trained on 771B tokens from 2.78B natural proteins. Unlike earlier protein LMs that only read sequences, it reasons jointly across sequence, structure, and function tracks, which makes it useful for generative protein design as well as understanding.

EvolutionaryScale (spun out of Meta's ESM team, now operating as Biohub) released smaller open-weights variants alongside the flagship, plus the ESM Cambrian family (300M–6B) as an efficient representation model. Access is via open weights on Hugging Face, the pip install esm SDK, or the hosted Biohub inference platform.

Specs

Context window

Not published as a token count — takes protein sequences and structural inputs; see docs for length limits per variant.

Pricing

Not published — hosted platform requires an account and API token.

Free quota

Open weights free to download and self-host; ESM Cambrian 300M/600M fully open.

API access

Biohub hosted API (biohub.ai), open-source Python SDK, Hugging Face weights

Weights

Open weights

Released

July 1, 2024

Benchmarks

Reported in the Science 2025 paper ("Simulating 500 million years of evolution with a language model"); evaluated on protein design and structure tasks — see the paper for task-level numbers.

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

Use cases

researchdata

Trust score — 9/10

/10 trustVerified

Scored 9/10 (verified tier, metrics-v1). Evidence: flagship protein foundation model; widely cited and built upon; actively maintained; org transitioned to Biohub with SDK continuity; open weights for smaller variants; flagship via hosted API. Manual review: the reference generative protein model; honest about variant-dependent access. License: Open weights for smaller variants; check Biohub terms for the flagship..

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
9.0/10
Openness · license + docs
8.0/10
Manual review · sanity read + red-flag check
9.0/10
  • · flagship protein foundation model; widely cited and built upon
  • · actively maintained; org transitioned to Biohub with SDK continuity
  • · open weights for smaller variants; flagship via hosted API
  • · manual review: the reference generative protein model; honest about variant-dependent access

Sources