Open weights vs open source AI: what the labels actually mean

"Open source AI" is the most overloaded phrase in the field. It gets applied to models you can download and run, models with published papers, and occasionally to a hosted API with a free tier. The distinctions matter, because they determine what your lawyers will let you ship.

Three different things people mean

Open weights. The trained parameters are published for download. You can run the model on your own hardware, inspect the numbers, fine-tune it, and serve it without calling anyone's API. Whether you can do so commercially depends entirely on the licence attached.

Open source (the actual standard). A licence meeting the Open Source Definition: freedom to use for any purpose, study, modify and redistribute, with no restrictions on who may use it or for what. Apache-2.0, MIT and BSD qualify. Custom licences that forbid certain uses do not, however reasonable those prohibitions may be.

Open research. A paper, and perhaps evaluation code, with no weights at all. Useful to read, impossible to deploy.

A model can be any combination of these. Most headline "open" releases are open weights with a custom licence and open research — not open source.

What custom licences typically add

The conditions are usually mild and occasionally decisive. Watch for:

• Acceptable-use policies. Lists of prohibited applications, incorporated into the licence by reference and revisable by the publisher.
• Scale thresholds. A clause requiring a separate agreement above some number of monthly active users. Irrelevant to a startup, potentially blocking for a large platform.
• Attribution and naming. Requirements to display "Built with X" or to prefix a derivative model's name with the base model's name.
• Derivative-licence propagation. Fine-tuned versions must ship under the same licence.
• Output restrictions. Clauses limiting use of the model's outputs to train competing models. These appear in both open-weights licences and hosted-API terms.

None of this is unusual in software licensing. The problem is only that the word "open" implies none of it applies.

Why the distinction is not pedantry

Three concrete situations where it bites:

White-labelling. If you resell a product built on a model whose licence requires naming the base model, your white-label offering is not actually unbranded.

Copyleft contamination. A small number of AI components ship under AGPL. In a hosted product, AGPL's network clause can require you to publish your own source. This is a licensing decision with product consequences, and it should be made deliberately.

Procurement. Enterprise and public-sector buyers increasingly ask for a software bill of materials that names every model and its licence. "It is open source" will not survive that review if the licence file says otherwise.

The questions worth asking before you build

1. Which licence file ships with the weights? Not the blog post — the file in the repository.
2. May I use it commercially, without a threshold? If there is a user cap, model your growth against it.
3. May I fine-tune and redistribute? Under what licence must the result be published?
4. May I use the outputs freely? Including to train or distil another model.
5. Does an acceptable-use policy apply, and can it change? A policy incorporated by reference can be revised after you deploy.
6. What does the model card say about training data? Not for legal comfort — there rarely is any — but so you know what you cannot claim about provenance.

What "open" buys you regardless of label

Even under a restrictive community licence, downloadable weights deliver things a hosted API cannot:

• Data residency. The text never leaves your infrastructure. For regulated sectors this is often the deciding factor.
• Version stability. The model does not change underneath you. Your evaluations stay valid.
• Cost structure. You trade per-token pricing for hardware you control — usually worse at low volume, better at high steady volume.
• Inspectability. You can probe activations, quantise, prune, and measure behaviour in ways an API forbids.

Those benefits are real and they are why open-weights releases matter, whatever the licence says. Just do not let the marketing word stand in for reading the terms — and when you write about a model, say "open weights" when that is what you mean.