Open weights are not
the whole source.

“Open” should tell you what you can do.

The Open Source Initiative’s Open Source AI Definition 1.0 describes an open-source AI system through four freedoms. A person must be able to use it for any purpose, study how it works, modify it, and share it with or without changes. Those freedoms apply to the system and to individual components such as model weights.

The definition is a practical reference for evaluating openness. A repository, a public API, or a downloadable file may be useful without granting all four freedoms.

Parameters are only one part of an AI system.

For machine-learning systems, the definition names three kinds of material needed to make meaningful changes: information about training data, the complete source code used to process data and train and run the model, and model parameters such as weights and configuration.

Publishing every raw training record is not always possible or required. The definition instead calls for detailed information about all training data—including data that cannot be shared—along with lists of public and third-party sources. That helps people understand the system and assess what can be reproduced without pretending every dataset is freely redistributable.

Under this definition, “open-source model” and “open-source weights” also require the data information and code used to derive those parameters. Weights alone can enable local inference or fine-tuning; they do not reveal the complete training process or establish the rights to reuse every component.

Read the Open Source AI Definition 1.0 ↗

Access and disclosure answer different questions.

A release can make weights available while saying little about how the model was trained, what data it used, or how it was evaluated. Another developer may publish detailed reports without granting the rights to modify or redistribute its model. Openness concerns both permissions and modifiable materials; transparency concerns what the developer discloses.

The 2025 Foundation Model Transparency Index assessed 13 developers using its own set of indicators. It found that open-weight developers scored higher on average than closed-weight developers, but also that openness did not guarantee transparency: several high-profile open-model developers still scored in the bottom half. The index measures developer transparency, not whether a particular release satisfies the Open Source AI Definition; its 2025 edition also changed some indicators from the prior year.

When comparing systems, keep the questions separate: Can you use and change it under the terms? Can you inspect how it was made? Can you reproduce the parts you need? A “yes” to one does not answer the others.

Explore the 2025 Foundation Model Transparency Index ↗

Check the release against what you need to do.

Before adopting a model, write down the job you need it to do. Then check the license and the specific files for that release—not just the headline or the repository description.

• Run it: are the parameters, inference code, supported hardware, and model version clear?
• Adapt it: do the license and training materials permit the changes you need?
• Audit it: are data sources, processing choices, evaluations, and limitations documented well enough for your purpose?
• Share it: do the code, parameters, and data information allow redistribution and derivative use on compatible terms?
• Deploy it: can you maintain the software dependencies and check model updates without losing track of the version you evaluated?

There is no single release choice that suits every project. A locally runnable model may matter most when you control the environment; permissive terms may matter most when you plan to redistribute; data provenance may matter most when a task has strict requirements. Make the decision against the work, the terms, and the evidence you can actually inspect.

For help comparing systems against a task, see how to choose an AI model ↗. For the distinction between model terms and generated content, read who owns AI-generated content ↗.