What do open-weight AI model licenses allow?

What is an open-weight model?

An open-weight model is one whose trained weights the developer lets the public download. The weights, also called parameters, are a large file of numbers; together with the code that runs them, they are the model.

That is all the term promises. It does not promise that you can use the model commercially, modify it, or build a competing product, and it does not promise that the training data or training code is available. Those questions are answered by the license, and downloading the weights does not by itself settle them.

This guide covers the private licenses on models. Government licensing is covered in AI licensing explained.

What does “open source” mean for AI?

The Open Source Initiative set a test for AI in October 2024: a model is open source only if users may use, study, modify, and share it, and can get its training code, its parameters, and detailed information about its training data.

For software, “open source” has had a fixed meaning for decades, maintained by the Open Source Initiative (OSI). A license qualifies only if it lets anyone use the software for any purpose, without discriminating against people or fields of use.

AI models raised a new question: whether releasing the weights is enough.

OSI answered on October 28, 2024, when it released the Open Source AI Definition 1.0 at the All Things Open conference. It requires four freedoms: to use the system for any purpose, study it, modify it, and share it.

It also sets a “preferred form” for modifying an AI system, with three parts:

• Data information: “Sufficiently detailed information about the data used to train the system,” so a skilled person could build a substantially equivalent system, including a list of publicly available training data and where to get it.
• Code: the complete source code used to train and run the system.
• Parameters: the weights and configuration settings.

The definition does not require releasing the training data itself. Data that cannot legally be shared must be described, not published.

Under this test, many models marketed as open are not open source. In July 2023, OSI said Meta’s Llama 2 license “is not Open Source,” because it restricts commercial use for some users and limits certain purposes through an acceptable use policy.

What do Apache 2.0 and MIT allow?

Apache 2.0 and MIT allow almost any use, including commercial use, provided you keep the required notices.

Apache 2.0 grants a worldwide, royalty-free, irrevocable copyright license to use, modify and distribute the work. It also grants a patent license from each contributor. That patent license ends for anyone who sues claiming the work infringes a patent. To redistribute, you must include the license, mark files you changed, and keep the original notices. It has no field-of-use restrictions.

MIT is shorter. You may use, copy, modify and sell the work, provided you keep the copyright notice and license text. The license text does not mention patents.

Several major model families now use these licenses:

• OpenAI’s gpt-oss models are listed under Apache 2.0 on their model page.
• Google moved Gemma 4, released April 2, 2026, to Apache 2.0, replacing the custom Gemma terms used for earlier versions.
• DeepSeek-R1’s model card says its code and weights are under MIT and expressly permits distillation.
• Meta released Muse Glimmer, a 30-billion-parameter model, under Apache 2.0 in August 2026, according to Thurrott, while keeping Muse Spark, which Thurrott describes as its most powerful model, closed.

There is one catch: the base model’s license can still matter. Apache 2.0 and MIT do not bind derivatives the way a custom license can, and DeepSeek’s own model card notes that its distilled versions built on Llama are derived from models originally licensed under the Llama 3.1 and Llama 3.3 licenses.

What does the Llama license require?

The Llama 4 Community License allows commercial use but adds conditions Apache 2.0 and MIT do not: a separate license from Meta for companies above 700 million monthly active users at release, Llama branding, and Meta’s acceptable use policy.

The agreement, effective April 5, 2025, is a custom license. Five of its conditions matter in practice:

• The 700 million user clause. If, on the Llama 4 release date, the products of you or your affiliates had more than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant at its sole discretion, and you have no rights under the agreement until it does.
• Attribution. If you distribute Llama or a product containing it, you must prominently display “Built with Llama.”
• Naming. A model you build with Llama and distribute must have “Llama” at the beginning of its name.
• Acceptable use policy. Your use must follow Meta’s policy, which bars uses such as weapons development, operation of critical infrastructure, fraud and disinformation.
• EU multimodal limit. For Llama 4’s multimodal models, the policy withholds the license from individuals domiciled in the EU and companies with a principal place of business there. End users of a product built on them are not covered by that limit.

The agreement is governed by California law.

For almost every business, the 700 million clause never applies. Its practical effect is to stop the largest technology companies from using Llama without a deal. The acceptable use policy, by contrast, applies to everyone.

What did the Gemma terms require before Gemma 4?

Before Gemma 4, Gemma models were released under Google’s own Gemma Terms of Use, which pass Google’s use restrictions down to every user and treat distilled models as derivatives. The terms were last modified April 1, 2026. They are still worth knowing: the terms’ appendix continues to list Gemma 1 through 3n, and Google’s Gemma 4 announcement cites more than 100,000 Gemma variants.

• Use must follow Google’s Prohibited Use Policy.
• Anyone redistributing Gemma or a derivative must make those restrictions enforceable on downstream users.
• Google reserves the right to restrict usage it reasonably believes violates the terms.
• “Model Derivatives” include models trained by distillation from Gemma or on synthetic data Gemma generated. Outputs alone are not derivatives.

The terms have no user cap. The distillation clause is the unusual part: it can reach a model that never copied a single Gemma weight.

What are RAIL and OpenRAIL licenses?

Responsible AI Licenses (RAIL) are the main attempt to combine open access with use restrictions.

The OpenRAIL-M license was announced on August 26, 2022, adapted from the license BigScience used for its BLOOM language model. It permits free access, reuse and redistribution. It attaches a list of restricted uses, and those restrictions must carry forward to derivative works. Redistributors may choose their own license as long as they include the use restrictions.

That carry-forward rule makes RAIL licenses resemble copyleft, but for behavior instead of source code. Because they restrict fields of use, they do not meet OSI’s definition of open source.

How do the main model licenses compare?

Apache 2.0 and MIT carry no use restrictions; the three custom licenses all add them.

License	Commercial use	Use restrictions	User cap	Attribution or naming	Express patent grant	Meets OSI definition?
Apache 2.0	Yes	None	None	Keep notices; mark changes	Yes	License qualifies; full system also needs code and data information
MIT	Yes	None	None	Keep notice	No	License qualifies; same caveat
Llama 4 Community License	Yes, below cap	Acceptable use policy; EU limit on multimodal models	700M monthly active users at release	“Built with Llama”; “Llama” name prefix	Check license text	No (OSI, 2023, on Llama 2)
Gemma Terms of Use (pre-Gemma 4)	Yes	Prohibited Use Policy, flowed down	None	Pass on terms; Notice file	Check license text	No
OpenRAIL-M	Yes	Listed restricted uses, flowed down	None	Include use restrictions	Check license text	No

Why does the license type matter under the law?

License type matters because the EU AI Act gives models released under a free and open-source license a lighter documentation load, and U.S. export rules controlled only closed weights.

The EU AI Act exempts general-purpose models released under a free and open-source license from two documentation duties, if the weights, the model architecture and information on model usage are public. The exception does not cover the copyright policy or the training-data summary. It does not apply at all to models with systemic risk, which the Act presumes for models trained with more than 10²⁵ operations. See frontier AI laws.

A license with use restrictions may not count as “free and open-source” for this purpose. Whether a custom license qualifies is a legal question the AI Office and courts have not fully settled as of October 2026.

U.S. export controls also draw the line at openness. The 2025 AI Diffusion Rule controlled only closed model weights. See AI export controls.

What should you check before building on an open model?

Before building on an open model, check eight things, starting with the license on the exact version you download:

• The license file on the exact model version you download, not the family name
• Any acceptable use or prohibited use policy incorporated by reference
• User or revenue caps
• Attribution and naming rules for products and derived models
• Whether distillation or synthetic data from the model is covered
• The license of any base model your chosen model was built on
• Where you and your users are located, for territorial limits
• Whether the license can change, and whether earlier versions stay valid

The data a model was trained on carries its own legal questions. See AI training data licensing.

Which way are model licenses heading?

Several major model families now ship under plain Apache 2.0, including Gemma 4, gpt-oss and Muse Glimmer. Custom licenses remain common, and their terms differ in ways that matter, so the license to read is the one on the exact model you download.

License terms quoted here are current as of October 2026; developers revise them. This guide is general information, not legal advice.