Open-weight vs open-source AI: what the label means and what labs actually release

People use open source, open weights and open model as if they meant the same thing. They do not, and the gap matters for developers, regulators and anyone reading a launch post. An open-weight model gives you the trained parameters. An open source AI system, under the Open Source Initiative's definition, gives you much more. This explainer sorts the terms using documents you can check.

What "open weights" means

Weights, or parameters, are the numbers a model learns in training. Releasing them lets anyone download the model and run it on their own hardware. Thinking Machines' model card for Inkling-Small, released July 30, 2026, is a typical example: open weights on Hugging Face in a standard and an NVFP4 version, under the Apache 2.0 license, with 276B total parameters, 12B active in a sparse mixture-of-experts design, and a context window up to 1M tokens.

The same card links a separate "Model Acceptable Use Policy," whose contents the card does not reproduce. So even a model with a permissive license can arrive with extra terms on how it may be used. Read both.

What "open source AI" requires

The Open Source Initiative's 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 output, and to share it, with or without modifications, for any purpose.

To make those freedoms real, the definition requires the "preferred form for modification" to be available, in three parts.

• Data information: "Sufficiently detailed information about the data used to train the system," enough for a skilled person to build a substantially equivalent system. This covers provenance, selection, labeling and processing, plus lists of public and third-party sources. It does not require the raw data itself to be shared.
• Code: "The complete source code used to train and run the system," including data processing and filtering, training, validation, testing, inference and architecture, under OSI-approved licenses.
• Parameters: The weights or other configuration settings, under OSI-approved terms, possibly including intermediate checkpoints and final optimizer state.

Weights under Apache 2.0 satisfy the third part. Whether a release satisfies the first two depends on what the lab publishes. A model card that lists parameters, a license and a context window does not, by itself, show the training code or data information.

Why the label is political

The distinction is not pedantry. In the United States, the fight over what to do about open models is live. According to TechCrunch, Nvidia's Jensen Huang shared on July 24, 2026 a policy letter signed by 25 companies, including Microsoft, Meta, Hugging Face, IBM and Palantir, urging the government not to over-restrict open-weight models. It said: "Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty." Note that the letter's own phrase is open-weight and open models, not open source.

On the other side, safety groups point to a hard fact about weights: once published, they cannot be recalled. The Paris Declaration on AI Safety and Security, signed November 9, 2026, acknowledges "the risk posed by the removal or circumvention of safeguards, including abliteration," while saying open-weight development should keep its benefits. Researchers have shown that refusal behavior can be removed from downloaded weights, a technique called abliteration.

A quick way to read a release

Ask five questions of any launch post.

• Are the weights downloadable, and in what formats?
• What license applies, and is there a separate acceptable use policy?
• Is the training code published, and does it cover data processing as well as training?
• Is there detailed information about the training data?
• Can you modify and redistribute the model for any purpose?

If the answer to the first is yes and the rest are unclear, you have an open-weight model. Only a release that clears the OSI test deserves the open source label.

What we do not know

The OSI page we checked does not give a date for version 1.0, and we have not audited any 2026 release against it line by line. We also cannot say how many frontier open-weight models publish training code or data information. Treat that as a question to put to each lab, not a settled fact. Our tracker on open-weight cyber capability shows how fast the gap between open and closed models is being measured.

A worked example from the news

Nvidia's own open-model push shows why the terms matter. TechCrunch lists a $6 billion deal in August 2026 between Nvidia and the coding startup Poolside to help build a US open AI ecosystem, and in October the expansion of Nvidia's Nemotron family of open-weight models with a 120-billion-parameter enterprise agent model and two lightweight Bolt models for PCs. Those are described as open-weight models. Whether any of them meets the OSI test depends on code and data information we have not reviewed, which is exactly the check this explainer recommends.

The Hugging Face platform sits in the middle of this. It hosts models, datasets and code under many licenses, and Nvidia's pending $12.93 billion purchase, covered in our reporting, has put its neutrality under regulatory questions. For builders, the lesson is to treat the license and use policy on each model page as the contract, not the marketing label.

Our take

Use open weights when you need control, privacy or low cost, and read the use policy as carefully as the license. Use the term open source only when the data information, code and parameters are all there. Regulators drafting rules for open models should say which one they mean, because a rule written for open source would barely touch most of what is released today.