Evolutionary Scale
Meta AI spin-out behind ESM3 — the protein language model that "simulated 500M years of evolution" — acquired by Chan Zuckerberg Biohub in late 2025.
1. Core Product / Service
Evolutionary Scale built ESM3, a multimodal generative protein model that reasons jointly over sequence, structure, and function, trained on billions of proteins (a 98B-parameter frontier variant plus smaller open releases) [1]. Its signature demonstration: generating esmGFP, a novel fluorescent protein at ~58% sequence identity to known GFPs — characterized as computationally simulating ~500M years of evolution [1]. Smaller variants (1.4B, 7B) were released open-weights for non-commercial research, with commercial access licensed to pharma via NVIDIA's BioNeMo [1].
Architecturally, ESM3 is a protein-specialized model, not a general LLM — this week's research stressed it carries no language-pretraining signal and instead represents sequence/structure, distinct from general-purpose models [local].
2. Target Users & Pain Points
- Pharma / biotech licensing commercial access for protein engineering and drug-discovery tasks.
- Academic researchers using the open non-commercial weights.
Pain solved: protein design traditionally requires costly directed evolution or screening; ESM3 generates function-conditioned sequences directly, collapsing search space.
3. Competitive Landscape
| Player | Model | Vs. Evolutionary Scale |
|---|---|---|
| alphafold3 | Structure prediction (Diffusion) | AF3 predicts structure; ESM3 also generates function |
| chai-discovery | Chai-1/Chai-2 open models | Both open-source-friendly; Chai targets antibody design |
| isomorphic-labs | AlphaFold lineage + wet lab | Isomorphic is the commercial wet-lab arm; ESM3 is a model/product |
4. Unique Observations
- The "Meta spinoff" arc: the team left Meta after the company dissolved its AI-protein group during the 2023 efficiency cuts, raised a record seed, shipped ESM3, then was acquired/acqui-hired by Chan Zuckerberg Biohub in Nov 2025 — a full cycle from big-tech research → venture startup → philanthropic research institute [1].
- General-LLM vs specialized-model debate: ESM3 is the clearest counterexample to the "just post-train a general LLM on protein data" thesis — this week's research found general LLMs (GPT-4o ~0.6, DeepSeek-R1 ~0.57 BERTScore) lag far behind protein-specialized models on protein tasks, with ESM3 as the canonical specialized model [local].
- Status change matters for the sector: its absorption into Biohub removes one of the few independent, well-capitalized protein-model companies from the commercial field, reinforcing the trend toward consolidation and big-tech/philanthropic capture.
5. Financials / Funding
- Seed: $142M (announced June 2024; led by Nat Friedman, Daniel Gross, Lux Capital; strategic from Amazon + NVIDIA NVentures) [1].
- Series A: $540M (July 2025) — total ~$682M at a reported ~$3.3B valuation [2].
- Exit: acquired/acqui-hired by Chan Zuckerberg Biohub (Nov 2025), ~50-person team; co-founder Alex Rives became CZI head of science [2].
6. People & Relationships
- Co-founders: Alex Rives (CEO), Tom Sercu, Salvatore "Sal" Candido (CTO) — ex-Meta AI/FAIR ESM team.
- Investors: Lux Capital, Nat Friedman, Daniel Gross, Amazon, NVIDIA (NVentures), Khosla Ventures.
- Acquirer: Chan Zuckerberg Biohub / Chan Zuckerberg Initiative.
- Peers: alphafold3, chai-discovery, isomorphic-labs.
Sources
- [1] EvolutionaryScale $142M launch coverage (2026-08-24)
- [2] AI Wiki — EvolutionaryScale profile (2026-08-24)