---
title: 'Open-source AI'
description: 'An AI system released with the rights and preferred form needed to use, study, modify, and share the system.'
canonical_url: 'https://darkfactory.dev/glossary/open-source-ai'
markdown_url: 'https://darkfactory.dev/glossary/open-source-ai.md'
collection: glossary
date_published: '2026-08-05T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Open-source AI


## Definition

Open-source AI is an AI system released with the rights and artifacts needed to use, study, modify, and share it. Under the Open Source Initiative's Open Source AI Definition 1.0, each freedom applies for any purpose and without requiring permission. The terms cannot discriminate against a person, group, or field of endeavor. The same requirements apply when a release describes a complete system, a model, its weights, or another structural element as open source.

The license is only part of the test. People also need the preferred form for making modifications. For a machine-learning system, OSI identifies three required classes of material:

- Data information must describe provenance, scope, selection, labeling, processing, filtering, and where available training data can be obtained. It must be detailed enough for a skilled person to build a substantially equivalent system.
- Code must cover data processing, training, validation, testing, architecture, and inference, including the settings needed to reproduce those procedures.
- Parameters must include the learned weights and other configuration required to run and modify the trained model.

The complete original training dataset does not always have to be redistributed. Unshareable data still has to be described, and publicly or commercially available data has to be listed with acquisition information. A checkpoint plus a vague model card does not meet this standard.

## Why it is contested

Industry usage is looser than the OSI standard. A downloadable model may be called open source even when its training code, data information, or redistribution rights are missing. Classification therefore requires separate answers to these questions:

1. Can the weights be downloaded?
2. Can they be used commercially or for any field of endeavor?
3. Can modified versions be redistributed?
4. Is the training and inference code available?
5. Is there enough information about the training data and process to study and reproduce the system?
6. Are essential components governed by compatible terms?

A release can be transparent in one respect and restrictive in another. "Open" remains useful when the speaker names the dimension, such as open weights or open data. This glossary reserves the unqualified open-source label for releases that meet the broader OSI threshold.

## Operational significance

Open-source AI can permit self-hosting, adaptation, offline operation, and independent testing. It can also reduce dependence on one API provider. Those are permissions and options, not guarantees that a system is affordable, secure, explainable, or reproducible on the operator's hardware.

Artifact access changes who carries operational responsibility. A hosted provider may handle patching, serving, and some abuse monitoring. A self-hosting organization inherits those jobs, along with provenance review, dependency management, data governance, evaluation, and incident response. Openness expands what an operator can control, but the operator still has to build the controls.

## Distinguish it from nearby terms

Open-weight means trained parameters are available under stated terms. It does not prove that the training code, data information, or modification rights satisfy an open-source standard. Source-available means some source or artifacts can be inspected, often under restrictions incompatible with open source. Free of charge describes price. Open access may describe an API while the model remains closed. None of those labels is a substitute for an artifact and license inventory.

## Check your understanding

A vendor releases weights and inference code under terms that permit research but prohibit commercial use. Training code and detailed data information are absent. Which parts are open, and why does the complete release fail the OSI test?

## Also called

open-source AI model, open-source model

## Related terms

- [Open-weight model](https://darkfactory.dev/glossary/open-weight-model)
- [Proprietary model](https://darkfactory.dev/glossary/proprietary-model)
- [Model weights](https://darkfactory.dev/glossary/weights)
- [Training data](https://darkfactory.dev/glossary/training-data)
- [AI system](https://darkfactory.dev/glossary/ai-system)

## Related factory areas

- [Model selection, routing & budgets](https://darkfactory.dev/factory/model-routing-budgets)
- [Security, privacy & software supply chain](https://darkfactory.dev/factory/security)
- [Runtime operations, observability, incidents & healing](https://darkfactory.dev/factory/runtime-operations)

## Evidence and further reading

- [Open Source Initiative: Open Source AI Definition 1.0](https://opensource.org/ai/open-source-ai-definition)
- [Stanford HAI Artificial Intelligence Glossary](https://hai.stanford.edu/ai-definitions)