ESM-2
MetaESM-2 is Meta's masked-language model family for protein sequences. The family spans several parameter sizes and produces representations that can be fine-tuned for protein prediction tasks; the Hugging Face release includes downloadable PyTorch, TensorFlow, and safetensors checkpoints and points to task-specific fine-tuning notebooks.
Openness
3 high confidence- weights
- open(MIT ungated PyTorch, TensorFlow, and safetensors files)
- data
- closed(card points to the paper rather than releasing the pretraining corpus)
- code
- partial(inference and fine-tuning integrations without pretraining pipeline)
- license
- MIT(OSI)
ESM-2 is an openly licensed and widely mirrored checkpoint family, but neither the training corpus nor a reproducible pretraining pipeline ships with this release.
- https://huggingface.co/api/models/facebook/esm2_t33_650M_UR50D recorded 2026-10-21
license: mit; gated: false; private: false; model.safetensors, pytorch_model.bin, and tf_model.h5 present.
- https://huggingface.co/facebook/esm2_t33_650M_UR50D/raw/main/README.md recorded 2026-10-21
Describes masked protein-language pretraining, points training-data detail to the paper, and links downstream fine-tuning notebooks rather than pretraining code.
Adoption
4 high confidenceAdoption is measured by Hugging Face downloads of the declared 650M checkpoint, one of several ESM-2 sizes Meta distributes; other checkpoint sizes are not counted here.
- https://huggingface.co/api/models/facebook/esm2_t33_650M_UR50D recorded 2026-09-10
downloads: 1274116.
Capability
2 high confidenceESM-2 documents one readable capability, an explicit fine-tuning pathway, two fewer than Prithvi-EO documents. Its broad downstream applicability across protein tasks does not convert into additional documented features.
- https://huggingface.co/facebook/esm2_t33_650M_UR50D/raw/main/README.md recorded 2026-09-10
Calls ESM-2 suitable for fine-tuning across protein-sequence tasks and links PyTorch and TensorFlow fine-tuning notebooks.
Verified 2026-09-10