Open-weight model licenses compared

Not legal advice. This page summarizes license texts as published on 14 November 2026, with short quotes. Licenses and use policies change, and how a clause applies depends on your facts and jurisdiction. For commercial or legal decisions, have a lawyer read the full terms.

Which license covers which model

The base models covered on this site fall under five licenses. The license fields below were checked on each base model’s Hugging Face repository on 14 November 2026.

Llama 3.3 70B Instruct and Gemma 3 27B are also gated on Hugging Face, so downloading them requires accepting the terms and waiting for manual approval.

Comparison: using the model

Comparison: redistributing and naming

Notes on each license

MIT: GLM-5.3-Flash and GLM-4.7-Flash

The GLM-5.3-Flash LICENSE file is the standard MIT text with “Copyright © 2026 Z.AI Co., Ltd”. It grants the rights “to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies” on one condition: “The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.” It has no use policy, no naming rule and no explicit patent clause.

Apache 2.0: Gemma 4, Qwen3.8 and gpt-oss-20b

Apache 2.0 adds more process than MIT. When you redistribute, Section 4 requires you to give recipients the license, mark modified files, keep existing notices and carry forward any NOTICE file. Section 3 grants a patent license that ends if you sue claiming the work infringes a patent. Section 6 grants no trademark rights beyond “reasonable and customary use in describing the origin of the Work”.

Google publishes the Gemma 4 license as the Apache License 2.0 text. Gemma 4 does not fall under the Gemma Terms of Use.

Llama 3.3 Community License

The Llama 3.3 license grants a “non-exclusive, worldwide, non-transferable and royalty-free limited license” to use, copy, modify and distribute. In return, Section 1.b requires a copy of the agreement, a “Built with Llama” display, the Notice text “Llama 3.3 is licensed under the Llama 3.3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.” and the naming rule. Section 2 sets the 700-million-user threshold. Section 5(a) grants use of the “Llama” mark only as needed for the naming rule.

The Llama 3.3 Acceptable Use Policy has a clause that matters for refusal-reduced models. Item 1 prohibits using Llama 3.3 to “Violate the law or others’ rights”, and its sub-item h names any action “to intentionally circumvent or remove usage restrictions or other safety measures, or to enable functionality disabled by Meta”. Whether making or using an abliterated Llama 3.3 falls under that clause is a legal question this page does not answer. It is a reason to read the policy with your own use case in mind.

The policy also withholds the Section 1(a) rights “with respect to any multimodal models included in Llama 3.3” from individuals and companies based in the European Union, though not from end users of products that incorporate them. Llama 3.3 70B Instruct accepts text only.

Llama 3.1 Community License

Apart from the version number and the policy it links to, the Llama 3.1 license text from Section 1 to Section 6 matched the 3.3 text word for word when we compared them. The Llama 3.1 Acceptable Use Policy is shorter. As published on 14 November 2026, it did not contain the 3.3 policy’s clause about circumventing safety measures. Both policies say the most recent copy is on Meta’s site, so check the current text rather than relying on this summary.

Gemma Terms of Use: Gemma 3

The Gemma Terms of Use apply to the models listed in their appendix: Gemma 1, 1.1, 2, 3 and 3n. Two definitions reach further than in other licenses. “Distribution” includes making Gemma available “as a hosted service via API, web access, or any other electronic or remote means”. “Model Derivatives” include models trained on Gemma’s synthetic outputs “to cause that model to perform similarly to Gemma”, though outputs themselves are not derivatives.

The Prohibited Use Policy lists “Attempts to override or circumvent safety filters” under dangerous, illegal or malicious activities. As with the Llama clause, how that applies to a refusal-reduced Gemma 3 variant is a question for your own legal review.

What changes with a third-party variant

A refusal-reduced variant is a modification of its base model. Abliteration edits the weights, a fine-tune trains them further, and a GGUF quantization converts them. The base license does not disappear in any of these steps.

• Using a variant brings in the base license’s use terms: the AUP or Prohibited Use Policy, the user threshold, and the law.
• Redistributing one, whether by uploading it, quantizing and re-uploading it, or serving it to others, brings in the redistribution duties in the table above.
• Permissive licenses let the variant author add terms, but only for their own changes. Under Apache 2.0, they may set different terms for their modifications provided the original work’s conditions are still met. The base notices stay.

When variant metadata disagrees with the base

A Hugging Face license tag is a one-word field filled in by whoever uploads the model. It is not a license file, and it can be wrong.

One example: the HauhauCS Gemma 4 26B A4B QAT variant names google/gemma-4-26B-A4B-it as its base model and carries the license tag gemma, which refers to the Gemma Terms of Use. Google publishes Gemma 4 under Apache 2.0, and the Gemma Terms of Use themselves say “For Gemma 4 terms, see the Gemma 4 license”. The card does not explain the tag.

When a variant’s metadata and its base disagree:

1. Follow base_model to the base repository and read its license file.
2. Look for a separate license file in the variant’s repository. A tag alone does not state terms.
3. Treat the disagreement as a reason to check, not as a new license. The base model’s license governs what the variant may be used for.
4. If it matters for your use, ask the variant’s author and record the answer.

For a walk-through of reading a variant’s card, see uncensored vs abliterated. For the difference between a permissive license and an open source model, see open-weight vs open source.

A checklist before you ship

1. Identify the base model and its version. Llama 3.1 and 3.3 have different use policies.
2. Read the base license file and every policy it incorporates by reference.
3. Check your use against the use policy, and your company’s size against any threshold.
4. If you redistribute, include the license, the required notices and any required display text.
5. If you publish a new model built from Llama 3.3 or 3.1, start its name with “Llama”.
6. If you serve a Gemma 3 model to others, include the use restrictions in your terms.
7. Save dated copies of the license and policy versions you relied on.

None of this replaces legal advice. It shows where the obligations are written, so you can ask a lawyer precise questions.