ESM3

ESM3, developed by EvolutionaryScale, is a frontier generative AI model designed to enable scientists to understand, imagine, and create proteins. Grounded in the principle that biology is fundamentally programmable — built from the same 20 amino acids across all living organisms — ESM3 comprehends the full breadth of biological data and serves as a powerful generative tool for protein science. It is intended for researchers and life sciences professionals working at the cutting edge of protein engineering, drug discovery, and synthetic biology.

ESM3 simultaneously reasons over the three fundamental properties of a protein — sequence, structure, and function — allowing users to input any combination of these data types as prompts and explore a vast space of protein possibilities. A companion model family, ESM Cambrian, defines a new frontier in protein sequence representation learning, delivering breakthrough performance and efficiency as a parallel offering to the generative ESM3 flagship.

Core Capabilities

• Simultaneous reasoning over protein sequence, structure, and function, enabling multi-modal prompting for generative protein design.
• Chain-of-thought prompting to guide the generation of novel proteins, as demonstrated by the creation of esmGFP, a novel Green Fluorescent Protein representing an evolutionary departure equivalent to approximately 500 million years of natural evolution.
• Support for diverse protein design tasks, including engineering proteins for applications such as carbon capture (carbonic anhydrase), plastic degradation (PETase), and antibody design.
• Representation learning via the ESM Cambrian model family, setting a new state of the art for understanding protein sequences.

Training Scale and Model Architecture

• Trained using 10^24 FLOPs — equivalent to 1 trillion teraflops of computing power.
• Trained on 2.78 billion natural proteins sampled from diverse organisms and biomes.
• Utilises 771 billion unique tokens of training data.
• Largest model contains 98 billion parameters, representing a true evolutionary-scale model.

Model Sizes and Availability

• ESM3 is available as a family of three model sizes: small, medium, and large.
• ESM3-open is a small but powerful model with weights and source code freely available on GitHub under a non-commercial license.
• All ESM3 models are accessible via the Forge API (currently in closed Beta, with access available by application).
• ESM3 is available on AWS via SageMaker, with planned availability on the AWS Omics platform and AWS Bedrock.
• ESM3 will be available on NVIDIA BioNemo and as an NVIDIA NIM microservice, optimised in collaboration with NVIDIA.

Responsible Development

• All models are built and deployed in accordance with EvolutionaryScale's responsible development framework.
• The company is committed to open, safe, and responsible AI research, scientific rigour, and advancing the frontiers of scientific discovery for the benefit of human health and society.
• Development is conducted in partnership with the broader scientific community.

EvolutionaryScale is joining forces with Biohub, further expanding its mission to develop AI that deepens biology's understanding for societal benefit. ESM3 is accessible through the Forge API, open-source GitHub repositories, and major cloud and AI infrastructure partners including AWS and NVIDIA.