Signum
Learn

Open-weight vs open-source AI: what the labels really mean

By Elliot Marsh, Editor · · 4 min read

Model launches love the word "open". But an "open" model might come with full training code and data, or it might be a file you can download under a licence that bans whole categories of use. The difference matters if you plan to build on it.

The short version

  • Open-weight means the trained model parameters (the weights) are published so you can download and run the model yourself. It says nothing on its own about the licence, the training data or the code.
  • Open-source AI, in the sense defined by the Open Source Initiative (OSI), means you get the freedoms to use, study, modify and share the system, which requires the weights, the code and enough information about the training data to rebuild something substantially equivalent, all under terms that do not restrict who can use it or for what.

Most "open" models today are open-weight. Far fewer meet the open-source definition.

What is in an AI model release

A trained model is more than one file. A full release could include:

  1. Weights: the billions of learned numbers. With these you can run the model.
  2. Inference code: the software to load and run it.
  3. Training code: how it was trained.
  4. Training data, or a detailed description of it.
  5. Documentation: a model card describing intended uses, limits and evaluations.
  6. A licence that sets what you may do with all of the above.

"Closed" models, such as those you reach only through an API, release none of the first four. Open-weight models release 1 and usually 2. Fully open releases include all of them.

The OSI definition

In October 2024 the Open Source Initiative, which maintains the long-standing Open Source Definition for software, published version 1.0 of the Open Source AI Definition. It requires that the system be made available under terms that grant the freedoms to use it for any purpose without asking permission, to study how it works, to modify it and to share it. To make study and modification possible, it requires access to the parameters, the full code used for training and running, and "sufficiently detailed information" about the training data. It does not require every piece of data itself to be published, a point that remains debated.

Licences: where the real differences hide

Two open-weight models can have very different terms.

  • Permissive licences such as Apache 2.0 and MIT let you use, modify and redistribute for any purpose, including commercially, with light conditions (keep the notice; Apache 2.0 also includes a patent grant).
  • Modified permissive licences. Some labs start from MIT or Apache and add conditions. Mistral Large 4, for instance, was reported to use a modified MIT licence. Always read the additions.
  • Custom "community" licences may restrict use above a certain number of users, ban specific uses, require attribution in product names or forbid using outputs to train competing models.
  • Research-only licences forbid commercial use entirely.

None of these restrictions are unusual or necessarily unreasonable, but under the OSI definition, limits on who may use a model or for what mean it is not open source.

Why the distinction matters

For businesses

The licence decides whether you can ship a product on the model, whether you must pay or ask permission past a certain size and whether a use you plan is banned. Getting it wrong is a legal risk.

For researchers

Without training code and data information, you cannot fully study why a model behaves as it does, check for contamination of benchmarks or reproduce results.

For governments and regulated sectors

Open weights allow self-hosting, which keeps data in-house and avoids dependence on one provider. That is a large part of the "sovereign AI" pitch made by European labs.

For safety

Open weights let anyone inspect and test a model, but also let anyone remove its safeguards. That trade-off is at the centre of policy debates about releasing the most capable models.

How to check a model in five minutes

  1. Find the licence file in the model's repository. Hugging Face shows a licence tag on each model page, but read the full text.
  2. Look for use restrictions: user thresholds, banned uses, rules on training other models with outputs.
  3. Check what is actually released: weights only, or training code and data details too?
  4. Read the model card for intended use, known limits and evaluation.
  5. Check dependencies: a fine-tuned model inherits the base model's licence terms.

A note on "open" washing

Because "open source" carries goodwill, it is sometimes used for models that are open-weight with restrictive licences. Neutral, precise phrasing helps: say "open-weight model under a custom licence" when that is what it is.

The bottom line

Open weights mean you can run the model. Open source means you can use, study, change and share the whole system freely. Many valuable models are the first without being the second. Before you build on any "open" model, read its licence.

Sources