Open weights is not open source

A model release announces itself as "open", you download the weights, and it runs on your own hardware. It is tempting to treat that the way you would treat an open-source library: free to use, modify and ship, forever.

Sometimes that is roughly true. Often it is not, and the difference only shows up when a lawyer reads the licence, which is usually after you have built something.

What "open weights" means

The weights are the trained parameters — the billions of numbers that make up the model. Publishing them lets you:

• run the model on your own hardware,
• fine-tune it on your own data,
• quantise it to fit a smaller machine,
• inspect its behaviour without an API between you and it.

That is a great deal, and for privacy, cost and reliability it can be decisive. But it is one component of what makes software open.

What open source traditionally requires

For ordinary software, "open source" has a specific meaning. The Open Source Definition requires that the source is available, that anyone can use it for any purpose including commercially, that modified versions can be distributed, and that the licence does not discriminate against people, groups or fields of use.

The weights of a model are closer to a compiled binary than to source code. The things you would need to rebuild the model — the training data, or at least a detailed description of it, and the training code — are usually not published.

In 2024 the Open Source Initiative published a definition specifically for AI systems. It asks for the parameters, the complete code used for training and running, and enough information about the training data that a skilled person could build a substantially equivalent system. Very few well-known models meet it. Most "open" releases are open weights, not open source in that sense.

Where the licences differ

This is the part that matters in practice. Model licences fall roughly into three groups.

Standard permissive licences. Some models are released under Apache 2.0 or MIT. These allow commercial use, modification and redistribution with minimal conditions — attribution, a copy of the licence, and in Apache's case a patent grant. If you are building a product, this is the easy case.

Custom community licences. Several of the most popular model families use their own licence written by the publisher. These often allow commercial use but add conditions that open-source licences do not have, for example:

• a cap on the size of company that may use it without a separate agreement, measured in monthly active users,
• an acceptable-use policy listing prohibited applications, which may be updated,
• naming requirements for derivative models,
• restrictions on using the model's output to train competing models.

Research or non-commercial licences. Weights are downloadable, but commercial use is not permitted at all.

The same publisher can use different licences for different releases. A model family's first version might be Apache 2.0 and the next under custom terms, or the reverse. Never assume one release's licence carries over to the next.

What else to check

The acceptable-use policy. If the licence incorporates a separate policy by reference, that policy is part of your obligations, and it can change.

Fine-tunes and merges. A fine-tuned model on a public hub inherits the base model's licence obligations, whatever its own model card says. A merge of several models may be bound by several licences at once.

Data provenance. Open weights say nothing about what the model was trained on. If your use case is sensitive to copyright or personal data questions, the lack of training-data information is a risk you are carrying, not the publisher.

Output restrictions. Some licences restrict using generated text to improve other models. If you plan to create synthetic training data, read that clause specifically.

A short routine before you build

1. Find the licence file in the model repository, not a summary on a blog.
2. Check whether it is a standard licence or a custom one.
3. If custom: read the commercial-use clause, any user-count threshold, and the acceptable-use policy it points to.
4. Check the licence of the exact version you are using, including any fine-tune.
5. Record what you checked and when, because licences and policies change.

Where AIonRadar helps

AIonRadar follows model releases with links back to the original announcements, model cards and repositories, which is where the licence actually lives. When a new release changes terms from the previous one, reading the source rather than a summary is what catches it.

Its comparisons and LLM API Selector cover the hosted side, and the GPU and VRAM calculator helps with the other half of the decision: whether an open-weights model you are allowed to use will also fit on the hardware you have.

It is free on the web, and the iPhone and iPad app has no account, no advertising and no in-app purchases.