In brief: Install the maintained esm package. You no longer need to clone the old Facebook ESM2 repository or use run_pretrained_model.py to start ESM3.
Choose between local inference and the Biohub API
ESM3 is a generative model for protein sequence, structure and function. Local inference offers direct control over a supported open checkpoint but needs substantial memory and compatible packages. The Biohub Platform provides a managed option with credentials and usage limits. Choose the approach based on your task, available compute, budget and data-handling requirements.
Do not confuse ESM3 with ESM2 or ESMFold2. Older commands such as esm2_t6_8M_UR50D refer to a different generation of models.
Create an isolated Python environment
Install a current supported Python release and create a clean environment before installing dependencies. On Windows, activate the virtual environment with the appropriate Scripts path; on Linux/macOS, use:
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install esm
python -c "import esm; print(esm.__file__)"Use the official README to confirm platform-specific PyTorch/CUDA dependencies. A successful Python import confirms installation, not that GPU inference or model downloads will work.
Connect to Biohub
Create a Biohub Platform account and API key. Store secrets in an environment variable or secret manager rather than embedding them in notebooks or public repositories. The following matches the official model quickstart; replace the placeholder with your own token:
import os
from esm.sdk.forge import ESM3ForgeInferenceClient
client = ESM3ForgeInferenceClient(
model="esm3-open-2024-03",
url="https://biohub.ai",
token=os.environ["ESM_API_KEY"],
)Model IDs and availability can change. If authentication fails, verify your account, token permissions, credits, network policy and the current API reference. This snippet creates a client; it does not yet run a biological inference.
Optional: run a model locally
See the official ESM3 README for the latest Hugging Face weights and model ID. Depending on package version, the model may be loaded using ESM3.from_pretrained(...). You may need Hugging Face authentication and enough GPU memory. Avoid assuming all ESM3 sizes can run on a laptop.
For researchers using shared clusters, pin the package version and document driver/CUDA versions. Never claim a run was reproducible without recording weights and environment details.
Troubleshooting checklist
- Import fails: confirm you installed
esmin the currently active environment. - CUDA not available: check
nvidia-smiand use the supported PyTorch build for your machine. - Out of memory: use smaller models or API inference; reduce sequence length and batch size as appropriate.
- API failure: check token, model ID, quota and documented endpoint.
- Unexpected outputs: remember that the ESM3 API works with structured
ESMProteinobjects, not generic ESM2 command-line scripts.
Primary sources and further reading
Editorial review: October 2026. This guide is educational. Model availability and APIs may change; confirm exact versions in the linked documentation. Generated structures and sequences require independent scientific validation.