Open-source licence

In short

An open-source licence is the legal permission that lets anyone use, study, change and share a piece of software, or an AI model, for any purpose. Common examples are the MIT, Apache 2.0 and GPL licences.

Also called: Open source license, OSI-approved licence, Open-source license
Topic: The AI business
Level: Beginner
First described: The term "open source" was coined at a strategy session in Palo Alto on 3 February 1998, and the OSI was founded later that month. Its Open Source Definition was derived from the Debian Free Software Guidelines¹⁰.

Software is protected by copyright by default. A licence is the document in which the owner grants other people rights to it. An open-source licence grants those rights broadly and to everyone.

The Open Source Initiative (OSI) keeps the most widely used test, the Open Source Definition. It has ten criteria. A licence must allow free redistribution, must come with source code, and "must not restrict anyone from making use of the program in a specific field of endeavor", such as business use¹. Licences that pass the OSI's review are listed as approved; MIT, Apache 2.0 and GPL versions 2 and 3 are among them².

Licences come in two broad kinds. Permissive ones, like MIT and Apache 2.0, let you fold the code into closed products. Copyleft ones, like the GPL, require that "all modified and extended versions of the program" stay free as well³.

AI made this harder, because a model is more than code. In October 2024 the OSI released the Open Source AI Definition 1.0⁴. It says an open-source AI system must let people use, study, modify and share it, and must come with the training and running code, the model parameters (weights), and "sufficiently detailed information about the data used to train the system"⁵.

That is why "open" models differ. OpenAI's gpt-oss-120b and Mistral 7B are published under Apache 2.0⁶⁷. Meta's Llama 3.1 licence, by contrast, requires any company with more than 700 million monthly active users to ask Meta for permission⁸, a restriction the Open Source Definition does not allow¹. Such models are better called open-weight.

Example

The UK government's Service Manual tells public bodies to publish new code under an OSI-compatible licence and notes that the Government Digital Service uses the MIT licence⁹. That is why much of GOV.UK's own code can be reused by anyone.

Why it matters

The licence, not the word "open" in a press release, decides what you may legally do with a model or program. Checking it matters before building a product on top of one.

Questions people ask

Is open source the same as free?

Not exactly. Open-source licences cannot charge royalties for redistribution¹, so the software usually costs nothing. But "free" in the free software sense is about freedom, not price³. Companies can still sell support, hosting or extra features around open-source code.

Is Llama open source?

Not under the OSI's definition. The Llama 3.1 licence requires companies with more than 700 million monthly active users to request a separate licence from Meta⁸, and the Open Source Definition bans licences that discriminate against people or groups¹. Llama is better described as an open-weight model.

What makes an AI model open source?

Under the OSI's Open Source AI Definition 1.0, released in October 2024, the model must come with its weights, the full code to train and run it, and enough detail about its training data for others to understand and rebuild it, all under terms that let anyone use, study, modify and share it⁴⁵.

What is the difference between MIT and GPL?

MIT is permissive: you can reuse the code in closed, paid products. The GPL is copyleft: if you distribute a modified version, it must stay free under the same terms³. Both are OSI-approved².

Related terms

Sources

Back to the glossary