Whether you may use an open AI model commercially is decided by the license file, not by the words "open source" on the model page. Apache 2.0 and MIT allow commercial use with no revenue cap. If you pass the model on, you essentially include the license text and the copyright notices. In October 2026 that covers, for example, Qwen3.8-27B, Gemma 4, gpt-oss, Muse Glimmer 30B and Mistral Small 4 under Apache 2.0, and DeepSeek V4 under MIT. The vendors' own licenses add conditions, and they differ a lot. Llama 4 does not grant the rights to its multimodal models to companies with their principal place of business in the EU. Mistral Medium 3.5 stops at 20 million USD monthly revenue, LFM2.5 at 10 million USD annual revenue. Several licenses require "Built with …" in the product, or a separate contract if you offer the model to others as a service. Teuken-7B-instruct v0.6 can't be used commercially at all.
I'm a developer, not a lawyer, and this isn't legal advice. It's my reading of the license texts, which I pulled myself on October 26, 2026. When in doubt, have your specific case checked. I read these texts because I build AI employees and test open models on my DGX Spark. Everything I write about local AI is collected on the AI page, and the models themselves with size and hardware needs are in Open Source AI Models 2026 Compared.
Open weight or open source: the difference
With its Open Source AI Definition 1.0, the Open Source Initiative sets out what "open source" means for AI. A system has to grant four freedoms: use it for any purpose "without having to ask for permission", study it, modify it, and share it. For that, the definition requires three components: sufficiently detailed information about the training data, the complete code for training and running the system, and the parameters, that is, the weights.
Open weight means you only get the third component. The weights are available for download, the training data and training code usually aren't. That applies to almost every well-known open model. For your business, though, the label matters less. What counts is what the license says about use, modification, redistribution and sale, and which conditions it attaches. That's exactly where the models differ widely.
Licenses of open models at a glance
As of October 26, 2026. The "Commercial" column is my reading for a small business that runs the model itself or builds it into its own product.
Apache 2.0 and MIT: the simple cases
Apache 2.0 grants every user a license that is "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable". So you may use, modify and redistribute the model permanently, worldwide and free of charge, including in a product you sell. The obligations mainly apply when you redistribute: "You must give any other recipients of the Work or Derivative Works a copy of this License", and modified files have to carry a prominent notice. Apache 2.0 explicitly grants no trademark rights. You may say your product runs on Qwen3.8-27B, but not pretend it's an Alibaba product.
MIT is even shorter. DeepSeek's license lets you use, modify, sell and redistribute the software "without restriction". The only condition: "The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software." For a small business running a model on its own hardware, both licenses are the simplest case.
Besides the license, check for a usage policy. Muse Glimmer 30B is under Apache 2.0, but its model card also points to a "Usage Policy" from Meta. It contains a list of prohibited uses, such as operating critical infrastructure, and the sentence "Muse Glimmer is not intended for individuals under the age of 18." With Apertus 1.5, you agree to the license and an "Acceptable Use Policy" before downloading. Read both texts as well if you build the model into a product for others.
Gemma: Gemma 4 under Apache 2.0, older versions not
With Gemma, the license depends on the generation. The Gemma 4 license page shows the Apache 2.0 text, and the model cards carry the "apache-2.0" tag. The older models fall under the Gemma Terms of Use. Google writes there: "The terms below apply to Gemma models listed in the Appendix at bottom of this page. For Gemma 4 terms, see the Gemma 4 license." The appendix lists Gemma 1 to 3 and Gemma 3n, among others.
The Gemma Terms also allow commercial use, but tie it to a prohibited-use list that you have to pass on contractually to your users when you redistribute. Google also reserves the right to restrict usage "remotely or otherwise" if it suspects a violation. If you still run a Gemma 3 model today, you're under different terms than with Gemma 4.
Llama 4: the EU clause
The Llama 4 Community License refers to a use policy that is "hereby incorporated by reference" into the license. That policy contains the sentence that matters for companies in Europe:
"With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union. This restriction does not apply to end users of a product or service that incorporates any such multimodal models."
Llama 4 Scout and Maverick process text and images, and Hugging Face lists both as image-text models. In my reading, a company with its principal place of business in the EU gets no rights under Section 1(a) to these models, so it may not run, modify or redistribute them itself. If you're an end customer using a product that contains Llama 4, the exception for end users probably applies. On top of that come the obligations from the license itself: display "Built with Llama" visibly, include a copy of the license, and name derived models with "Llama" at the start. Above 700 million monthly users you need a separate license from Meta.
In the license text of Muse Spark Open, Meta's new open model from October 15, 2026, I found no such EU clause. There, Meta names Meta Platforms Ireland as the contracting party for Europe. The Muse Spark Open usage policy is attached to the license text as "Exhibit A" and is therefore part of the terms. What's required is the "Built with Muse Spark Open" notice in every product, and anyone giving others access to the model as a service needs a contract with Meta above 20 million USD revenue in twelve months. The text explicitly exempts internal use and features within your own product.
Qwen: two licenses in one family
Within Qwen 3.8, the license depends on the model. Qwen3.8-27B, the model on my DGX Spark, is under Apache 2.0. Qwen3.8-Flash-Next and the large Qwen3.8-2.4T-A95B have their own licenses.
A note on Qwen3.8-Flash-Next: the Qwen Community License 1.0 requires a separate license from Qwen before commercial use if you or an affiliate run "a Model as a Service or AI Work Assistant business". According to the text, an "AI Work Assistant" is a standalone AI product primarily designed for AI-assisted coding or office productivity. Internal use is exempt, as long as neither the model nor its outputs become available to third parties. For companies without such a business, commercial use is allowed. Above 100 million monthly users or 20 million USD monthly revenue, the model name has to be displayed visibly in the product, and the same applies to Qwen3.8-Max. I don't use Flash-Next for client projects and therefore haven't had the clause reviewed further. If you plan to build a product for office work or coding on Flash-Next, clarify this first.
Revenue thresholds and "Built with": Mistral, Kimi, MiniMax, Liquid
Several licenses tie the rights to the size of your company:
- Mistral Medium 3.5: "You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million", measured on the previous month. For most small businesses that's no obstacle. The smaller Mistral models such as Mistral Small 4 and Ministral 3 are under Apache 2.0 anyway.
- Liquid AI's LFM2.5: commercial use is only licensed as long as your annual revenue stays below 10 million USD.
- Kimi K3: a contract with Moonshot AI is only needed if you offer the model to others as a service and have more than 20 million USD revenue in twelve months. Above 100 million monthly users or 20 million USD monthly revenue, "Kimi K3" has to be displayed visibly in the product. Internal use is exempt.
- MiniMax M3: any commercial use requires the "Built with MiniMax M3" notice and a one-time email notice to MiniMax, and prior written approval above 20 million USD annual revenue from the product.
- GLM-5.3: only providers of models as a service with more than 10 billion USD revenue need a security review by Z.ai.
Teuken: open, but not commercial
Teuken-7B-instruct v0.6 from the German OpenGPT-X project was trained in all 24 official EU languages. The model card names CC BY-NC 4.0 as the license and describes the purpose like this: "designed for private, non-commercial, research, and educational use". That rules this version out for client work. Hugging Face also hosts an older version called Teuken-7B-instruct-commercial-v0.4, which according to Hugging Face is under Apache 2.0. I haven't tested either.
What the license doesn't cover: privacy
A license says what you may do with the weights. It says nothing about where your data ends up. If you run an open model on your own hardware, your input doesn't leave your network, whether the model comes from the US, China or France. If you use the same model through the vendor's app or API, or as a cloud model in Ollama, that provider's privacy terms apply. What can still leak in a local setup is covered in Local AI and Privacy: Cloud or Server?, and the cloud side, using Claude as the example, in Claude Privacy and GDPR: What Applies. I show the difference between app, API and your own weights using DeepSeek as the example in Running DeepSeek Locally: What V4 Needs.
How to check a model license in five steps
- Open the license file, don't read the label. On Hugging Face, custom licenses often just say "other". The text sits in the repository as
LICENSE or in the access form.
- Place your use case. Do you use the model internally, build it into a feature of your product, or offer it to others as a service? Many clauses only apply in the third case.
- Check the thresholds. Monthly or annual revenue, user numbers, with or without affiliates.
- Note the obligations. License copy, "Built with …", naming rules, notifying the vendor.
- Record the date. Licenses change with new versions. Keep the text with its retrieval date and check it again with every model update.
Frequently asked questions
Can I use Apache 2.0 models for client projects?
Yes. Apache 2.0 allows commercial use with no revenue cap. If you pass the model on, you include the license text and mark your changes. In October 2026 that applies to Qwen3.8-27B, Gemma 4, gpt-oss, Muse Glimmer 30B and Mistral Small 4, for example.
Can a company in Germany use Llama 4?
According to the use policy, companies with their principal place of business in the EU get no rights under Section 1(a) of the license to the multimodal Llama 4 models. Llama 4 Scout and Maverick are multimodal. End users of a product that contains such a model are exempt. That's my reading, not legal advice.
Can DeepSeek be used commercially?
The weights of DeepSeek V4 and V4.1 are under the MIT license. Commercial use is allowed, the condition is the copyright and license notice. Privacy is a separate question: run locally, no data goes to DeepSeek; through the app or API, it does.
What does "Model as a Service" mean in a license?
It means giving third parties access to the model itself, for example through an API, so that they control inputs and parameters. The Kimi K3 and GLM-5.3 licenses exclude end-user products where the model only sits inside specific features. Muse Spark Open also exempts features within your own product.
Can Teuken be used commercially?
Teuken-7B-instruct v0.6 can't, it's under CC BY-NC 4.0. The older Teuken-7B-instruct-commercial-v0.4 carries the Apache 2.0 tag on Hugging Face.
Does the license also cover the text the model generates?
The licenses handle that differently. OpenMDW-1.1 states explicitly that there are no obligations for outputs. Meta claims no ownership of outputs from Muse Spark Open, and Google claims no rights in outputs in the Gemma Terms of Use either. For other licenses, read the text.
Where you go from here
Before you pick a model, check the table with size and hardware needs in Open Source AI Models 2026 Compared. How an open model becomes a building block of an AI employee, and which other routes exist, I compare in Building an AI Employee: Six Ways Compared. What the hardware costs is in Your Own AI Server: Hardware and Cost.
