MLSTAC

A single, consistent way to publish and consume machine learning models, built on the STAC MLM extension and Safetensors.

Get started Quickstart

Experimental

The API may still change between minor versions.

---

Why MLSTAC

• Models as metadata

  ---

  Every model is a STAC Item described with the MLM extension. The weights travel as Safetensors, so the description and the data stay in sync.

• Many backends, one call

  ---

  Load from HTTP(S), local disk, Amazon S3 or Google Cloud Storage. mlstac.load figures out how to reach the model for you.

• Metadata first, weights later

  ---

  Inspect a model before downloading a single byte. Pull the files only when you are ready, straight from the loader.

• Ensembles out of the box

  ---

  Point at a list of .pt2 files and MLSTAC builds an ad-hoc ensemble with a minimal STAC description on the fly.

---

Install

pip install mlstac

In 30 seconds

import mlstac

# 1. Load only the metadata (no weights yet)
model = mlstac.load("https://example.com/my-model/mlm.json")

# 2. Look before you leap
print(model.get_model_summary())
model.print_schema()

# 3. Download, then build a usable model
net = model.download("./my-model").compiled_model()

---

Explore the docs

• Getting Started

  ---

  Install MLSTAC and run your first load.

  Installation

• User Guide

  ---

  Loading from any source, and downloading the right way.

  Loading models

• API Reference

  ---

  Every class and function, generated from the source.

  ModelLoader