ESM-3
B+EvolutionaryScale (Meta origin)
Protein Language ModelSequence GenerationFunction Predictionopen
Updated 3 weeks agoNextIn take
ESM-3 is the GPT-4 moment for protein language models — genuinely multimodal and shockingly capable — but the compute requirements mean most teams will use the smaller variants and miss the best performance.
Specifications
| Architecture | Transformer-based protein language model with multimodal conditioning |
| Parameters | ~98B |
| Training Data | UniRef, MGnify, PDB structures, functional annotations — 2.78B protein sequences |
| License | Open weights under EvolutionaryScale license, free for research and commercial use. |
| Hardware | 8x A100 80GB for full model; smaller variants available |
| Inference Cost | Self-hosted or via EvolutionaryScale API (pricing varies) |
| API Available | Yes |
| Weights Available | Yes |
Benchmark Performance
| Benchmark | Score |
|---|---|
| ProteinGym | 0.47 Spearman (avg) |
When to Use This
- +Variant effect prediction and protein fitness landscapes
- +Zero-shot protein function annotation
When NOT to Use This
- −Structure prediction — use dedicated structure models instead
- −Resource-constrained environments without GPU access
Production Readiness
research
Known Users
- EvolutionaryScale partners
- Academic labs with compute access
Grade Rationale
B+
The most capable open protein language model by a wide margin, with genuine multimodal understanding of sequence, structure, and function. B+ because the 98B parameter model is expensive to run, and the smaller open variants sacrifice significant capability.
Sources
Update History
2026-03-15Initial entry