Choosing a model | BioLM Docs

The BioLM catalog has dozens of models — encoders, structure predictors, generators — and more are added over time. This guide is about the step before you write a call: figuring out which model fits your task and confirming the exact actions, input type, and params it accepts. Once you know that, the call itself follows the single pattern from What are BioLMs?.

There are three ways to discover models — the web catalog, the CLI, and the Python API. They all draw from the same source, so pick whichever fits your workflow: browse visually, script against the terminal, or stay in a notebook.

Start from the task, not the model

Most searches start with a task. Map the task to one of the four inference actions, then look for a model that advertises that action:

Embeddings / similarity / features → the encode action. Reach for an encoder such as esm2-8m (proteins) and use the vectors for search, clustering, or as inputs to a downstream model.
Structure or property prediction → the predict action. esmfold predicts 3D structure from a sequence; other predictors return property scores or classifications.
New sequences (de novo or conditioned) → the generate action. progen2-oas generates antibody-like sequences; inverse-folding models like ProteinMPNN generate sequences for a target backbone.
Reference / precomputed data → the lookup action, which takes a query= instead of items=.

A model only exposes the actions it supports, so “which model?” and “which action?” are really one question. The tools below all report the supported actions next to each model.

The web catalog

The model catalog is the fastest way to browse. Each model has a page listing its supported actions, expected input type, accepted params, and the request/response schemas — plus copy-pasteable snippets. Use it when you are exploring or want the authoritative per-model reference; this documentation intentionally does not duplicate it.

From the command line

The biolm model command group turns discovery into a scriptable loop:

bash

# List the catalog; filter, sort, and export for scripting
biolm model list
biolm model list --filter encoder=true --sort model_name
biolm model list --format json --output models.json

# Inspect one model: actions, input type, and full JSON schemas
biolm model show esmfold --include-schemas

# Print a copy-pasteable example for a model and action
biolm model example progen2-oas --action generate

biolm model list accepts --filter (e.g. encoder=true, model_name=esm2), --sort (prefix a field with - for descending), and --format (table, json, yaml, or csv, optionally written to a file with -o). biolm model show prints a model’s metadata; add --include-schemas to see the exact request and response shapes for every action. biolm model example emits ready-to-run code — in Python, Markdown, RST, or JSON — so you can go from “which model?” to a working call in one step.

See biolm model for the complete command reference.

From Python

The same discovery lives in the SDK, which is handy inside notebooks and scripts:

python

from biolm import list_models, get_example, Model

# The full catalog as a list of dicts (name, slug, supported actions, ...)
models = list_models()

# A copy-pasteable example for a model and action
print(get_example("esmfold", action="predict"))

# Or ask a Model instance directly
model = Model("progen2-oas")
print(model.get_example(action="generate"))   # one action
print(model.get_examples())                    # every supported action

list_models() returns the catalog so you can filter it however you like in Python. get_example() (and the equivalent Model.get_example / Model.get_examples methods) generates usage snippets; all of them accept a format= argument ("python", "markdown", "rst", or "json"). This means you can confirm a model’s action and required arguments and get working code without leaving your session.

Discovering a local catalog

If you route inference through a self-hosted biolm-hub gateway rather than biolm.ai, point the CLI and SDK at it with biolm hub set (defaulting to http://127.0.0.1:8000). Afterward, biolm model list, show, and example — and the Python discovery helpers — resolve against the hub’s catalog, so you discover exactly the models your gateway serves. You can also browse them in a browser at http://127.0.0.1:8000/catalog while the gateway is running. See Running inference through BioLM Hub for setup details.

Where to go next

What are BioLMs? — the four actions and the call pattern, with examples.
Running an inference — run a model from Python or the CLI.
Working with biological data — load sequences and structures from local files.
Model finetuning — train models on your labeled data.
Running inference through BioLM Hub — route inference through a local or self-hosted gateway.
Quickstart — install, authenticate, and make your first call.
biolm model — complete biolm model command reference.
biolm hub — route discovery and inference through a local gateway.
biolm.models — the full Model API, including get_example.
BioLM model catalog — per-model actions, input types, params, and request/response schemas.