Skip to main content
Back

Chan Zuckerberg Biohub Republishes Meta's ESM-2 650M Protein AI Model

Chan Zuckerberg Biohub is now hosting the ESM-2 650M protein language model on Hugging Face, continuing stewardship of the ESM lineage after Meta's team spun out.

  • ESM-2 650M protein language model now mirrored at biohub/esm2_t33_650M_UR50D on Hugging Face.
  • Chan Zuckerberg Biohub is the active steward of the ESM package line including ESM C and ESMFold2.
  • Checkpoint is 0.7B params, 33 layers, MIT licensed, Safetensors and PyTorch formats.
  • Over 3.1M downloads and ~4K likes on the mirror; 39 finetunes tracked.
  • NVIDIA ships a TransformerEngine version with FP8 on Hopper and Blackwell GPUs.
  • Community Flash Attention port claims up to 70% faster inference and 60% less memory.

Chan Zuckerberg Biohub has published the 650-million-parameter ESM-2 protein language model in the Biohub repository on Hugging Face. The listing traces the model to Meta FAIR’s original ESM-2 release and provides a new model ID for loading the checkpoint through Transformers and other Hub-based tooling.

ESM-2 is a transformer trained to infer masked amino acids from surrounding sequence context. That masked-token objective produces representations that researchers reuse for protein classification, residue-level prediction, contact estimation, and variant scoring. The 650M release has 33 transformer layers and belongs to the UR50D checkpoint family described in the ESM-2 paper.

A new model ID changes the plumbing

Changing from facebook/esm2_t33_650M_UR50D to biohub/esm2_t33_650M_UR50D affects dependency manifests, local caches, access controls, deployment configuration, and artifact provenance. Applications that pin the Meta model ID will continue requesting that repository until their configuration changes.

The architecture and claimed weight lineage match the original release. Production teams requiring exact continuity should still compare the tokenizer, configuration, tensor names, shapes, values, and selected artifact hashes before substituting one repository for another. A matching model card or architecture name does not establish byte-for-byte identity.

Property Biohub listing Operational consequence
Model ID biohub/esm2_t33_650M_UR50D Requires an explicit configuration change
Architecture ESM-2 masked language model Loads through standard Transformers ESM classes
Scale Approximately 650 million parameters across 33 layers Memory use depends on precision, sequence length, and batch size
Artifacts PyTorch, TensorFlow, and Safetensors formats