Protein Models API - Biohub Platform

The Protein Models API runs inference on ESMC, ESMFold2, and ESM3. For larger or longer-running workloads, use the Batch API.

Install the esm package

To run esm locally, download the PyPI package. You can also call the API directly over HTTP without it, but the package provides a simpler way to access the API and additional useful libraries.

pip install esm

To begin using the esm package, create an API key and connect to the model you want to use:

from esm.sdk.forge import ESMCForgeInferenceClient

client = ESMCForgeInferenceClient(model="esmc-6b-2024-12", url="https://biohub.ai", token="<your API token>")

For ESMFold2 use SequenceStructureForgeInferenceClient, and for ESM3 use ESM3ForgeInferenceClient. See Get started for each model and Tutorials for code snippets that demonstrate how to use the esm SDK.

Operations

Operation	Models	Purpose
fold	ESM3, ESMFold2	Fold proteins. Defaults to `esmfold2-fast-2026-05`.
fold_all_atom	ESMFold2	Fold molecular complexes containing proteins, DNA, RNA, and ligands. Defaults to `esmfold2-fast-2026-05`.
logits	ESM3, ESMC	Run one inference step and return logits, embeddings, and SAE features (ESMC only).
encode	ESM3, ESMC	Tokenize sequence, structure, or function annotations.
decode	ESM3, ESMC	Decode sequence, structure, or function annotation tokens.
generate	ESM3	Generate an output track conditioned on inputs.
generate_tensor	ESM3	Generate an output track from raw tokens and return raw tokens.
forward_and_sample	ESM3	Run one inference step and sample output tokens.
inverse_fold	ESM3	Inverse fold proteins. Defaults to `esm3-open-2024-03`.

Example: send a request

Send a JSON body with Content-Type: application/json and authenticate with your API key:

curl https://biohub.ai/api/v1/fold \
  -H "Authorization: Bearer $BIOHUB_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "esmfold2-fast-2026-05",
    "sequence": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQ"
  }'

See each operation’s page for its request fields and response schema.

Response formats

Responses are JSON by default. Most operations also support Python pickle responses, which can reduce payload size for tensors and arrays. The esm Python SDK requests pickle automatically. See Response formats.

Input limits

On fold and fold_all_atom, ESMFold2 models accept inputs up to 768 tokens. For inputs up to 4096 tokens, or to run many predictions asynchronously, use Batch folding.

On-demand or batch

Use	When
Protein Models API	You need a result in the same request, for interactive or low-volume work.
Batch API	You have many inputs, long inputs, or long-running jobs such as maximum-accuracy folding and binder design.

Protein Models

Batch API

ESM Atlas