ESM3 by EvolutionaryScale: status and timeline

ESM3

A generative language model creating novel proteins by reasoning simultaneously about sequence, structure and biological function. Trained on 2.78 billion natural proteins.

DeployStage 4 of 5

API available via AWS and NVIDIA; open-weight version released under MIT.

Updated 25 Jun 2024Checked 10 Oct0 updates this week

Milestones

Next · Biohub commits to 10,000 GPUs by 2028

ESM3 unveiled with $142M seed roundJun 2024Complete.
esmGFP: ESM3-designed fluorescent proteinJun 2024Complete.
AWS and NVIDIA partnership announced2024Complete.
Acquired by CZI to form BiohubNov 2025Complete.
Biohub commits to 10,000 GPUs by 2028Target 2026Current milestone.

Most important updates

Upcoming

2026Biohub commits to 10,000 GPUs by 2028 (next)

Current obstacles

Designs must work in living systemsLab designs must fold inside cells, avoid triggering immunity, and function in complex environments.
Scaling to therapeuticsMoving from research tool to drug requires clinical validation and regulatory approval.

Physics limits

Training data captures explored space onlyESM3 learns from billions of known proteins but underrepresents extreme designs, limiting extrapolation far from examples.
Protein folding remains hardSmall amino acid changes can cause misfolding, aggregation or toxicity. Models predict structure imperfectly.

How it works

3 parts

Multimodal input

Mix constraints from any layer

Scientists provide partial information: structure, motif, or function. ESM3 completes by reasoning over all three.

Generative reasoning

Chain-of-thought protein design

ESM3 generates completions token-by-token, learning which sequences fold stably and produce desired functions.

Massive scale

Learn from the landscape

Training on 2.78 billion proteins gives detailed knowledge of how mutations, structures and functions relate.

Spec sheet

Spec	ESM3
Training compute	10^24 flops^R (reported)
Training set	2.78 billion proteins^R (reported)
Model sizes available	Small, medium, large via API^R (reported)

^R reported by the company

Papers & demos

1 item

Jun 2024demo
esmGFP: de novo design of a fluorescent protein (opens science.org)ESM3 designed a novel functional protein from scratch, announced June 2024 with publication in Science January 2025.

Update log

2 updates

Tue 25 Jun 2024

Major: EvolutionaryScale unveils ESM3, raises $142M seed from Friedman and GrossPressHardware

Jun 2024

Major: ESM3 designs esmGFP, a novel fluorescent protein from scratchPaperResearch

About EvolutionaryScale

The team behind ESM3

EvolutionaryScale

ESM3 protein language model, AI for biology

EvolutionaryScale builds AI models that understand, imagine and create proteins. Founded by researchers from Meta's FAIR lab. In November 2025, acquired by the Chan Zuckerberg Initiative to form Biohub.

ESM3 reasons over protein sequence, structure and function in a generative model trained on 2.78 billion natural proteins.
Designed esmGFP, a novel fluorescent protein compressing ~500M years of natural evolution into computation.
Acquired by CZI in Nov 2025 to form Biohub, focusing on curing all diseases by 2100 through AI and biology.

evolutionaryscale.ai (opens evolutionaryscale.ai)Company profile →

Founded: 20233 yrs
Headquarters: United States
Status: Acquired
Valuation: $142MprivateJun 2024
Raised: $142M1 round
Last round: Seed · $142MJun 2024
Works in: AIAI for scienceBiotech
Coverage: 1 program · 3 updateslatest 1 Nov 2025checked 10 Oct
Lead investors: Nat FriedmanDaniel Gross
Partners: Amazon Web ServicesNVIDIA
Acquired by: Chan Zuckerberg Initiative (opens biohub.org)
People: Alexander RivesCEO and Chief ScientistTom SercuCo-founderSal CandidoCo-founder

Spec

ESM3

Training compute

10^24 flops^R (reported)

Training set

2.78 billion proteins^R (reported)

Model sizes available

Small, medium, large via API^R (reported)