BioNeMo ESM-2 650M
Protein / BiologyNVIDIA · BioNeMo · v1.0 · released
About
ESM-2 650M is a 33-layer transformer trained as a protein language model on ~65M UniRef50 sequences. NVIDIA BioNeMo provides containerized inference (NIM) and training-optimized FlashAttention kernels for the model.
Intended use: Per-residue protein embeddings for structure prediction (downstream models like ESMFold), variant-effect prediction, evolutionary search, drug-target identification. Self-host via NeMo or NIM container.
Architecture
- Type
- encoder
- Parameters
- 650M
- Layers
- 33
- Hidden dim
- 1,280
Protein language model — masked-language transformer over amino-acid sequences (UniRef50 corpus). Vocabulary is 33 tokens (20 canonical amino acids + special tokens). Produces per-residue contextual embeddings used for structure prediction, variant-effect prediction, and embedding-based search. Originally published by Meta (ESM-2); BioNeMo packages it with NVIDIA-optimized training + inference for drug-discovery workflows.
Memory
- Weights (BF16)
- 1.30 GB
- Weights (FP8)
- 0.65 GB
- Activation estimate
- 0.80 GB
Pricing
Free — open weights
Self-host on your own GPU. The calculator surfaces GPU-hours cost on the hardware page instead of an API price.
Provenance
- Source
- docs.nvidia.com
- License
- mit
- Hugging Face
- facebook/esm2_t33_650M_UR50D
- Last verified
- 2026-06-25