# Open model licenses: what Apache 2.0, MIT, Llama and Mistral let you do

> An open model's license says what you may do with its weights. Apache 2.0 and MIT allow commercial use, changes and sharing if you keep the notices. Meta's Llama 4 license adds a "Built with Llama" notice, a naming rule and a 700 million user limit. Mistral's license asks businesses that host the model as a service for a separate license.

Mindket editors · Published 2026-11-06 · Updated 2026-11-06 · https://mindket.com/learn/open-weight-models/open-model-licenses

## In short

- Put five questions to any license file before you build on a model: sell, change, share, credit and host.
- Pass the license text and notices along with anything you redistribute.
- Read the license of the exact version you download, since a family can change its terms between releases.
- Have a lawyer read the file before you ship a product on it, because a lesson is not legal advice.

## Before you start

You need a model page open in your browser and a place to write answers. A license is a contract, and this lesson is a guide to reading one, not legal advice. A company that ships a product on a model should get legal advice on the file first.

Each model keeps its license in up to three places, and each answers a different question. The card's settings block names it. The full text sits behind a link or in a file named LICENSE inside the repository. Some makers add a separate use policy.

## Read a model's license in five questions

Hugging Face's [docs](https://huggingface.co/docs/hub/model-cards) say the card's settings block holds a license field, which the Hub shows on the model page and lets people filter by. Common values such as `apache-2.0` and `mit` name a standard text. A custom license is set as `other`, with a name and a link to its text, as in this example from the docs. When the text has no URL, the docs say the card can link to a LICENSE file stored in the repository.

````yaml
# Example from https://huggingface.co/coqui/XTTS-v1
---
license: other
license_name: coqui-public-model-license
license_link: https://coqui.ai/cpml
---
````

1. Open the model page and read the `license` value in the card's settings block. A value such as `apache-2.0` or `mit` names a standard text. `other` marks a custom license. Look for a `license_name` and a `license_link` beside it, or open the Files tab and look for a file named LICENSE.
2. Open the LICENSE file on the Files and versions tab. Search it for "commercial", "monthly active users" and "service".
3. Look for a use policy next to it. The gpt-oss-20b repository has a USAGE_POLICY file beside its Apache license, the Llama 4 Scout repository lists a USE_POLICY.md beside its LICENSE, and the Llama 4 license brings in Meta's Acceptable Use Policy by reference.
4. Answer five questions in a note. May you sell something built on it? May you change or fine-tune it? May you share your changed copy? What notice must you show? May you host it for other people?
5. Add the license name, the model version and the date you read it. Hugging Face's license list holds a separate value for every Llama version, so a new version can bring new terms.

## What the four licenses allow

The Apache and MIT licenses are the short ones. GitHub's [summary of the Apache license](https://github.com/github/choosealicense.com/blob/gh-pages/_licenses/apache-2.0.txt) lists commercial use, modification, distribution and patent use as permitted, with two conditions: keep the copyright and license notices, and document your changes. Its [MIT summary](https://github.com/github/choosealicense.com/blob/gh-pages/_licenses/mit.txt) drops the patent permission and the change log, leaving one condition, the copyright notice. Neither forces you to release your changes under the same license, a rule called copyleft. gpt-oss's card puts it as building "without copyleft restrictions or patent risk".

Meta's [Llama 4 Community License](https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE) is longer. If you share Llama materials, or a product that contains them, you pass on a copy of the agreement. You also show "Built with Llama" on a related website, interface, blog post, about page or product documentation. If you use Llama materials or outputs to train or improve an AI model that you make available, its name must start with "Llama". If your products or your affiliates' had more than 700 million monthly active users in the month before the release date, you must request a license from Meta, which it may refuse. The license ends if you start litigation against Meta that alleges Llama infringes your intellectual property.

Mistral's license on [Mistral Large 4](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE) calls itself a modified version of the MIT license. It lets you use, copy, modify, run, deploy, further train and sell copies of the model. The one added condition concerns hosting.

> the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose.
>
> Mistral, Section 2 of the license, for businesses that operate a Model as a Service, [Mistral Ecosystem Open-Weight License](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE) (read 2026-11-06)

**What each license asks of you**

| | Apache 2.0 | MIT | Llama 4 Community License | Mistral Ecosystem Open-Weight License |
| --- | --- | --- | --- | --- |
| Commercial use without asking | ✓ Yes | ✓ Yes | Up to 700 million monthly users | Yes, unless you run a Model as a Service business |
| Change and fine-tune it | ✓ Yes | ✓ Yes | ✓ Yes | ✓ Yes |
| What you must keep or show | The notices, and a record of your changes | The copyright notice | The agreement, "Built with Llama" and a Notice file | The copyright and permission notice |
| Rule on naming a model you train from it | ✗ No | ✗ No | The name starts with "Llama" | ✗ No |
| Use policy built into the license | ✗ No | ✗ No | ✓ Yes | ✗ No |

Sources: [GitHub: choosealicense, Apache License 2.0](https://github.com/github/choosealicense.com/blob/gh-pages/_licenses/apache-2.0.txt) (read 2026-11-06), [GitHub: choosealicense, MIT License](https://github.com/github/choosealicense.com/blob/gh-pages/_licenses/mit.txt) (read 2026-11-06), [GitHub: Llama 4 Community License Agreement](https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE) (read 2026-11-06), [Hugging Face: Mistral Ecosystem Open-Weight License](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE) (read 2026-11-06)

## Where Mistral draws the hosting line

The license defines a Model as a Service business as giving a third party access to the model or its derivatives for inference or post-training, through an API or hosted endpoints. That party must be able to meaningfully control the inputs, parameters or training data for it to count. The definition leaves out end-user products with the model's abilities built into specific features, and mere relaying of requests to models hosted by others.

It also exempts internal use, meaning any use that does not make the model or its derivatives available to third parties, and any use through Mistral's own products. A team that runs the model for its own staff is outside the condition. A company that sells API access to it is the case the definition names.

## Where six models keep their license

The six cards read on 6 November 2026 differ in what the license field says and in where the text sits. Llama 4 Scout's card sets the field to `other` and names the license `llama4` in a separate field. Gemma 4's card links Google's own license page, which this lesson did not read, so read that page before you rely on the card's label.

| Model | License field | Where the text is |
| --- | --- | --- |
| gpt-oss-20b | apache-2.0 | A LICENSE file, plus a USAGE_POLICY file |
| Qwen3.8-27B | apache-2.0 | A LICENSE file |
| DeepSeek-V4.1-Flash | mit | A LICENSE file, and the card says the repository and the model weights are under the MIT License |
| Gemma 4 12B | apache-2.0 | No file in the repository, a link to Google's license page in the card |
| Llama 4 Scout | other, with the license name llama4 | A LICENSE file and a USE_POLICY.md in the gated Hugging Face repository, and Meta's llama-models repository on GitHub |
| Mistral Large 4 | other | A LICENSE file with the Mistral Ecosystem Open-Weight License |

## Common mistakes

- Stopping at the license badge. gpt-oss-20b carries a standard open license and still ships a usage policy file.
- Assuming a new version keeps the old terms. Hugging Face lists one value per Llama version, and Gemma 4's card names an Apache license while the same list still holds a value for the Gemma Terms of Use.
- Dropping the notice when you redistribute. Every license in this lesson asks you to pass a notice along.
- Trusting a re-upload's license field. A quantized or fine-tuned copy has its own card, so check its `base_model` field and read the original model's license too.
- Reading the Llama license without its use policy. The license includes the Acceptable Use Policy by reference, and the policy is a separate document.

## Try it yourself

- Open the LICENSE files of Qwen3.8-27B and Mistral Large 4 and answer the five questions for each in a note.
- Search the Llama 4 license for "monthly active users" and write down the number and the date it counts from.
- Find a quantized copy of a model you like on Hugging Face and check which original model its card names.
- Decide whether your own plan for a model, in one sentence, sits inside or outside Mistral's Model as a Service definition.

## Next: choose an open model

A license narrows the list of models you can use. [Lesson 3](https://mindket.com/learn/open-weight-models/choose-an-open-model) narrows it again by the job and by the computer you have.

## Check yourself

1. What does the value "other" in a Hugging Face card's license field tell you?

   A. The model has no license at all
   B. The license is a custom one, and the card points to its text
   C. The model is for research only
   D. The license is Apache 2.0

   Answer: B. Hugging Face's docs say a custom license is set as `other`, with a name and a link to the text. When the text has no URL, the card can link to a LICENSE file in the repository.

2. Under the Llama 4 Community License, when must you request a license from Meta?

   A. When you fine-tune the model
   B. When you sell a product built on it
   C. When your products had over 700 million monthly active users in the month before the release date
   D. When you host it on a cloud service

   Answer: C. Section 2 of the license sets that threshold, and Meta may grant or refuse the request.

3. What does Mistral's license add to the MIT license?

   A. A separate commercial license for businesses that operate a Model as a Service
   B. A ban on commercial use
   C. A rule that derived models start with "Mistral"
   D. A "Built with Mistral" notice

   Answer: A. The one added condition is the Model as a Service rule, with an exemption for internal use.

4. What does Apache 2.0 ask when you share a copy?

   A. Name the copy after the maker
   B. Ask the maker first
   C. Share your changes under the same license
   D. Keep the copyright and license notices and document your changes

   Answer: D. GitHub's summary lists those two conditions, and says Apache 2.0 does not force copyleft on your changes.

## Frequently asked questions

### Can I use an open-weight model in a paid product?

Often yes. The permissive licenses ask only for notices, Meta asks for its own license above a user threshold, and Mistral asks for one from companies that sell the model as a service. Read the exact file for the version you use.

### What is the difference between Apache 2.0 and MIT?

Both let you use, change and sell the software. GitHub's license summaries list a patent grant and a duty to document changes only for Apache 2.0. MIT is shorter and asks only that you keep the copyright and license notice.

### What happens if I break a license?

Under the Llama 4 license, Meta may terminate the agreement for a breach, and you must then delete the Llama materials and stop using them. Mistral's text has no termination clause, and it provides the model "as is", without warranty.

### Can I fine-tune an open model and sell the result?

Usually. All four licenses in this lesson grant the right to modify the model or make derivative works. Check the extra conditions first, which are the naming rule in the Llama license and the hosting rule in Mistral's.

### What does "license: other" mean on a Hugging Face card?

It marks a custom license outside Hugging Face's list. Hugging Face's docs say to add a name and a link to the text, or a LICENSE file in the repository. Mistral Large 4's card uses it with a LICENSE file, which holds the Mistral Ecosystem Open-Weight License.

## Sources

- [GitHub: choosealicense, Apache License 2.0](https://github.com/github/choosealicense.com/blob/gh-pages/_licenses/apache-2.0.txt) (read 2026-11-06)
- [GitHub: choosealicense, MIT License](https://github.com/github/choosealicense.com/blob/gh-pages/_licenses/mit.txt) (read 2026-11-06)
- [GitHub: Llama 4 Community License Agreement](https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE) (read 2026-11-06)
- [Hugging Face: Mistral Ecosystem Open-Weight License (LICENSE file)](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE) (read 2026-11-06)
- [Hugging Face: mistralai/Mistral-Large-4-1T-A52B model card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B) (read 2026-11-06)
- [Hugging Face: openai/gpt-oss-20b model card](https://huggingface.co/openai/gpt-oss-20b) (read 2026-11-06)
- [Hugging Face: meta-llama/Llama-4-Scout-17B-16E-Instruct model page (gated)](https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct) (read 2026-11-06)
- [Hugging Face API: meta-llama/Llama-4-Scout-17B-16E-Instruct (license field and file list)](https://huggingface.co/api/models/meta-llama/Llama-4-Scout-17B-16E-Instruct) (read 2026-11-06)
- [GitHub: openai/gpt-oss README](https://github.com/openai/gpt-oss) (read 2026-11-06)
- [Hugging Face API: openai/gpt-oss-20b (file list)](https://huggingface.co/api/models/openai/gpt-oss-20b) (read 2026-11-06)
- [Hugging Face: Qwen/Qwen3.8-27B model card](https://huggingface.co/Qwen/Qwen3.8-27B) (read 2026-11-06)
- [Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) (read 2026-11-06)
- [Hugging Face: google/gemma-4-12B-it model card](https://huggingface.co/google/gemma-4-12B-it) (read 2026-11-06)
- [GitHub: Meta Llama 4 model card](https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md) (read 2026-11-06)
- [Hugging Face Docs: Licenses](https://huggingface.co/docs/hub/repositories-licenses) (read 2026-11-06)
- [Hugging Face Docs: Model Cards](https://huggingface.co/docs/hub/model-cards) (read 2026-11-06)

[Browse kits](https://mindket.com/kits)