Open weights vs open source AI: what the difference means and how to check a model

An open-weight AI model lets you download its trained parameters. An open source AI system, under the Open Source Initiative's definition, also gives you detailed information about the training data and the complete code used to train and run it. Most models called open today are open-weight only.

The difference is not pedantry. It decides what you can legally do, what you can inspect and who controls the model once you build on it.

What open weights give you

Open weights means the numerical parameters of the trained model are published. You can run it on your own hardware, fine-tune it and often host it for customers. You usually cannot see the data it was trained on or reproduce the training run.

Recent releases show the range. Meta's Muse Spark Open, released October 15, 2026, has 816 billion total parameters and ships under the Muse Model License, which Meta describes as having limited commercial terms. Nvidia's Nemotron 3.5 line is reported by VentureBeat to use the OpenMDW license, which permits commercial use, fine-tuning and redistribution, and Nvidia says it released training recipes and datasets. Mistral's Mixtral 8x7B went out under Apache 2.0, an established open source software license.

Those are three quite different bargains, all loosely called open.

What the Open Source Initiative requires

The Open Source AI Definition 1.0 sets a higher bar. An open source AI system must grant four freedoms: to use the system for any purpose "and without having to ask for permission," to study how it works and inspect its components, to modify it for any purpose including changing its outputs, and to share it with or without modifications.

To make those freedoms real, the definition requires access to the preferred form for making modifications, in three parts:

• Data information: "Sufficiently detailed information about the data used to train the system," including its origin, how it was collected, labeled, processed and filtered, and a list of public training datasets.
• Code: "The complete source code used to train and run the system," covering data processing, training, validation, inference and the model architecture.
• Parameters: "The model parameters, such as weights or other configuration settings."

Note what is not required: the training data itself. The definition asks for information about the data, including data that cannot be shared. But weights without that information and code do not qualify.

Where common models land

By that test, a release with weights and a restrictive custom license is open-weight, not open source. A license that limits commercial use or requires permission fails the first freedom. A release with weights but no data information or training code fails the second.

We have not audited every model against the definition, and the answer for any specific model depends on the license text and what is actually published. The point is that the question has a checkable answer, and "open" in a press release does not supply it.

Meta's own framing is explicit about the trade-off. Alexandr Wang, Meta's chief AI officer, said, as reported by The Next Web: "Because of both safety and capabilities, we will generally not open-source our most powerful models." The strongest models stay closed, and the weights that go out are a tier below.

Why regulators care about the label

The EU AI Act treats free and open models differently. As The Next Web summarizes it, Article 53 exempts genuinely free and open models from some documentation duties, but the exemption does not cover systemic-risk models. A model marketed as open but released under a restrictive license may not get the exemption, and a very capable one may not either way. That is a question for the European Commission, which can fine up to 3 percent of global turnover.

What it means in practice

• If you are building a product: read the license before the model card. Check commercial limits, user-count thresholds and acceptable-use clauses, then plan for the license changing.
• If you need auditability: weights alone will not let you study training data. Look for the data information the OSI definition requires.
• If you worry about control: downloaded weights keep working even if a vendor changes its API, which is a real advantage over a hosted model, and the main reason open weights matter regardless of the label.
• If you care about safety: open weights cannot be recalled. Meta says it added extra safety post-training for that reason.

For a recent example of a tiered release, see our report on Mistral Large 4 and its open weights.

A checklist for reading any launch

Before you call a release open, run it through a short list. Is the license text on a standard, recognized license, or a custom one with usage conditions? Are weights downloadable without an approval step? Is there a model card describing training data sources, even at a high level? Is training code published, or only inference code? Are there thresholds, such as user counts or revenue, that switch the license terms? A yes to the first four and a no to the last is close to what the Open Source Initiative describes. Most launches today answer yes to the second and little else.

Our take

Use "open weights" for downloadable parameters and save "open source" for systems that meet the OSI definition. The distinction costs nothing and stops marketing from doing the thinking. When a lab says its model is open, ask three things: what can I download, what license binds me and what can I verify about how it was made.