Open-weight vs open-source AI

Two separate questions

“Open-weight” and “open source” answer different questions. The first is practical: can you download the trained parameters and run them yourself? The second is about rights and materials: do the terms let anyone use, study, modify and share the model, and did the developer publish what someone would need to rebuild it?

Almost every model on this site is open-weight. The files are on Hugging Face and run on your own hardware. That tells you nothing about the license. GLM-5.3-Flash ships under MIT, Gemma 4 under Apache 2.0, Llama 3.3 under Meta’s own community license and Gemma 3 under Google’s Gemma Terms of Use. These licenses allow different things, and none of them, on its own, makes a model open source in the OSI sense.

What the OSI definition requires

The Open Source Initiative published version 1.0 of the Open Source AI Definition (OSAID). It starts from four freedoms. An Open Source AI must be available under terms that let anyone:

• Use the system “for any purpose and without having to ask for permission”;
• Study how it works and inspect its components;
• Modify it for any purpose, “including to change its output”;
• Share it with others, with or without modifications, for any purpose.

The definition then says what you need in hand to exercise those freedoms, which it calls the “preferred form to make modifications”. For a machine-learning system, that form must include all three of these elements.

Three details are easy to miss.

First, the data itself does not have to be published. The definition allows “unshareable data” as long as it is described. OSI’s FAQ gives medical data and personally identifiable information as examples of data that cannot always be shared.

Second, weights alone cannot qualify. The OSAID states that “Open Source models” and “Open Source weights” “must include the data information and code used to derive those parameters.”

Third, copyleft is allowed. The terms may require that modified versions be released under the same terms as the original.

OSI also reported which models passed its validation phase: Pythia, OLMo, Amber, CrystalCoder and T5. It named Llama 2, Grok, Phi-2 and Mixtral among the models analyzed that did not pass, and it adds that these results “are not certifications of any kind”. None of the base models covered on this site appears on either list, and we have not assessed any of them against the definition.

Where open-weight releases stop

OSI’s explainer on open weights describes the gap: a weights release does not include the training code, the training dataset or, when the data cannot legally be shared, “comprehensive data transparency”. You can fine-tune, quantize or edit the weights. You cannot reproduce how they were made.

This matters for refusal-reduced models. An abliterated or fine-tuned variant changes weights whose training data nobody outside the original developer can inspect. You can measure how the variant behaves. You cannot trace why the base model behaved as it did.

What each license lets you do with the weights

The license is a separate layer from the OSAID’s materials test. The table summarizes what the license texts say for the four licenses that cover most models on this site. It paraphrases; read the full text before relying on any row.

MIT and Apache 2.0 both appear on OSI’s list of approved licenses. Both were written for software, and OSI’s FAQ notes that the definition “does not take a stance on the legal nature of Parameters”.

For Llama, OSI wrote in 2023 that Meta’s Llama 2 license “is not Open Source”. It pointed to two things: the commercial limit for some users and the use restrictions in the Acceptable Use Policy. The Llama 3.3 license keeps both, in Section 2 and Section 1.b.iv.

The Gemma Terms of Use have no user threshold. They do incorporate a Prohibited Use Policy, and Google “reserves the right to restrict (remotely or otherwise) usage” that it reasonably believes breaches the terms.

A decision list for your use case

Work through these questions in order, for the base model of whatever you plan to run. A refusal-reduced variant does not change the answers, because its weights are a modification of the base model’s weights.

1. Which license governs? Find the base_model on the variant’s card and open the base repository’s license. Do not rely on the variant’s license tag alone; our license comparison shows a case where the two disagree.
2. Do you need approval to download? Llama 3.3 and Gemma 3 are gated on Hugging Face and need manual approval (checked 14 November 2026).
3. Is your use on a restricted list? MIT and Apache 2.0 name no restricted purposes; the law still applies. For Llama 3.3 and Gemma 3, read the use policy against your actual use case.
4. Is your company very large? Only the Llama license has a threshold: more than 700 million monthly active users on the Llama 3.3 release date.
5. Will you share the weights or a modified copy? Uploading a quantized file or a fine-tune is redistribution. Each license then has the notice duties listed in the table.
6. Will you serve the model to other people? Gemma’s terms treat a hosted service as distribution. Llama’s terms attach their attribution duties to anything you “make available”. The MIT and Apache 2.0 texts do not mention hosting.
7. Will you publish a new model? Under Llama 3.3, if you use Llama materials or their outputs to create, train or improve a model that you distribute or make available, its name must start with “Llama”.
8. Do you want to call it open source? Then check all three OSAID elements, not only the license on the weights. If the developer has not published training code and data information, the accurate description is “open-weight”, followed by the license name.

This list describes license texts, not how they apply to your facts. It is not legal advice. For anything with commercial or legal stakes, have a lawyer read the full terms against your situation.

Common misconceptions

“If I can download it, it is open source.” Downloadable weights make a model open-weight. The OSAID also asks for data information and code, and for terms that allow any purpose.

“Apache 2.0 weights make it open source AI.” An OSI-approved license on the parameters covers one of three elements. The other two, data information and code, are separate requirements.

“Open source means no obligations.” MIT requires the copyright and permission notice in “all copies or substantial portions of the Software”. Apache 2.0 requires the license text, prominent notices on changed files and the original attribution notices when you redistribute.

“All Gemma models share one license.” They do not. The Gemma Terms of Use list the models they cover (Gemma 1, 1.1, 2, 3 and 3n) and say “For Gemma 4 terms, see the Gemma 4 license”, which Google publishes as the Apache License 2.0. A refusal-reduced Gemma 4 variant and a Gemma 3 variant therefore come with different terms.

“A refusal-reduced variant comes with a fresh license.” Editing weights does not remove the base terms. An abliterated Llama 3.3 70B still carries the Llama 3.3 license and its Acceptable Use Policy.

“Nobody knows if weights are copyrighted, so the license does not bind me.” OSI’s FAQ says the legal status of parameters “will become clearer over time”. Do not read that uncertainty as permission. Meta’s and Google’s terms both say that using the materials means you agree to be bound by them.

How we describe models on this site

Our editorial policy uses “open-weight” whenever the weights can be downloaded, then names the license. We use “open source” only for a model that meets the OSAID, or inside a quote. In practice that gives descriptions like these:

• GLM-5.3-Flash: open-weight, MIT license.
• Gemma 4 26B A4B: open-weight, Apache 2.0.
• Llama 3.3 70B Instruct: open-weight, Llama 3.3 Community License, gated download.
• Gemma 3 27B: open-weight, Gemma Terms of Use, gated download.

The same discipline applies to the word “uncensored”, which tells you even less than “open-weight”. What uncensored AI means covers that label.