What Is Open-Source AI?

How open-source AI is defined

Open source started with software. It means the source code is published under a license that lets anyone use it for any purpose, see how it works, change it, and pass it on, with or without changes. AI is harder to pin down, because a working language model is more than code. It also depends on its weights (the values learned during training) and on the data it was trained on.

In October 2024, the Open Source Initiative (OSI), the nonprofit that maintains the widely used Open Source Definition for software, published version 1.0 of its Open Source AI Definition. It requires four freedoms (to use, study, modify, and share the system) and says a qualifying release must include:

• The weights, so you can run and adapt the model.
• The code used to train and run it.
• Detailed information about the training data, enough for a skilled person to build a substantially equivalent system, including where any public or third-party training data can be obtained. The definition does not require publishing all of the data itself.

Not everyone agrees with this definition, and many people still use “open-source AI” loosely. The definition may also be revised over time, so check OSI’s website for the current version.

How to tell whether a model is really open source

Many models marketed as open share only their weights. Under the OSI definition, those are open-weight models, not open-source AI, and the open-weight model entry compares the two side by side. When a vendor or developer calls a model open source, these questions show what you are actually getting:

• What license does it use, and does that license allow your intended use, including commercial use?
• Can you download the weights, or is the model only available through someone else’s service?
• Is the training code published?
• Is there documentation describing the training data and where it came from?

A model can be useful and well supported without being open source. The point is to know which trade-offs you are accepting before you build on it.

More than models: the open-source AI toolkit

Open-source AI is not only about models. Much of the software used to build AI systems is open source too, and businesses often rely on it even when the model itself is proprietary. Common examples include:

• Machine learning frameworks such as PyTorch, used to train and run models.
• Model libraries such as Hugging Face Transformers, which make published models easier to download and use.
• Local runners that let you run a model on a laptop or your own server.
• Building blocks for AI apps, such as open-source vector databases and frameworks for connecting models to documents, tools, and workflows.

Each component has its own license. A chatbot might combine an open-source framework, an open-source database, and a proprietary model reached over the internet, so “open” can describe parts of a system without describing the whole thing. When you evaluate an AI product, it is reasonable to ask which parts are open, which are not, and who maintains each one.

What full openness adds for a business

Many benefits people associate with open AI come from having the weights at all: running a model in your own environment, keeping data in-house, adapting it through fine-tuning, and not losing access when a vendor retires a version. Those apply to open-weight models too and are covered in that entry. A release that meets the stricter open-source standard adds a few things:

• Insight into training data. Documentation of where the data came from can help when you need to assess bias, copyright, or privacy exposure, or answer a customer’s due-diligence questions.
• Fewer strings attached. Open-source terms let you use the model for any purpose, without the restrictions on certain uses or on very large companies that some open-weight licenses carry. Attribution and similar conditions can still apply.
• The ability to audit or rebuild. With the training code and data details, a skilled team can check how the model was made and, with enough data and computing power, train a comparable one if the original project stalls.

The trade-offs match any model you run yourself: you pay for servers and for the computing power each request uses (see inference), you handle security and updates, support usually comes from a community or a paid partner, and quality varies, so test candidates on your own tasks first.

Further reading

• AI ImplementationUseful if you want help connecting systems and deploying AI once you have weighed open-source and hosted options.