ESM-3

weights

open(ungated esm3_sm_open_v1.pth with structure and function decoder weights alongside)

data

closed(2.78 billion natural proteins described by count, corpus not released)

code

partial(inference, generation and tutorial code in the esm package

license

MIT(OSI

The open checkpoint moved namespaces and lost its license metadata on the way - the EvolutionaryScale repo now redirects to biohub and the destination card has no license field at all. MIT is still stated, in the card body and in the repository LICENSE.md under a Chan Zuckerberg Biohub copyright, so the grant exists; confidence is medium because it is asserted in prose rather than declared on the artifact. Only the 1.4B tier is distributed; the 7B and 98B models are served through the platform API and are outside this axis.

• https://huggingface.co/api/models/EvolutionaryScale/esm3-sm-open-v1 recorded 2026-09-10

  Redirects to huggingface.co/api/models/biohub/esm3-sm-open-v1; gated: false; private: false; no license tag and no cardData license key; data/weights/esm3_sm_open_v1.pth, esm3_structure_decoder_v0.pth and esm3_function_decoder_v0.pth present.

• https://huggingface.co/biohub/esm3-sm-open-v1/raw/main/README.md recorded 2026-09-10

  Front matter declares only language and tags, no license. The body says "This repository is under a MIT license" and links the esm repository LICENSE.md; training is described as 2.78 billion natural proteins, with no corpus released.

• https://raw.githubusercontent.com/Biohub/esm/main/LICENSE.md recorded 2026-09-10

  "License (MIT)", Copyright 2026 Chan Zuckerberg Biohub, Inc., with the standard MIT permission grant.

• https://raw.githubusercontent.com/Biohub/esm/main/_assets/ESM3_README.md recorded 2026-09-10

  "These models are available under the MIT license." Code covers local inference and generation and a tutorials cookbook; the 7B and 98B tiers are reachable only through the hosted platform.

2,772 downloads in the trailing 30 days for biohub/esm3-sm-open-v1, which is the openly distributed checkpoint and what this band covers. The larger tiers are served through the hosted Forge API and publish no artifact count, so they are outside the counted scope rather than folded into it.

• https://huggingface.co/api/models/biohub/esm3-sm-open-v1 recorded 2026-09-10

  downloads: 2772; likes: 330.

Two attributes read straight off the publisher's own description. The third does not follow from them - the release documents how to prompt and sample from the model, not how to adapt it, and the fine-tuning guidance in the repository is written for ESMC rather than ESM-3. All three are read off the openly distributed 1.4B checkpoint and its documentation, which is the same SKU the openness axis scores; the hosted 7B and 98B tiers are not read here, and 'no second scale' is a statement about what this publisher distributes rather than about what it serves.

• https://raw.githubusercontent.com/Biohub/esm/main/_assets/ESM3_README.md recorded 2026-09-10

  "able to jointly reason across three fundamental biological properties of proteins: sequence, structure, and function ... represented as tracks of discrete tokens at the input and output"; "ESM3 is a generative masked language model. You can prompt it with partial sequence, structure, and function keywords, and iteratively sample masked positions".

• https://huggingface.co/biohub/esm3-sm-open-v1/raw/main/README.md recorded 2026-09-10

  "a generative model capable of designing proteins conditioned on partial prompts of sequence, structure and function"; usage points to the esm repository notebooks, with no fine-tuning entry point.