“Open” AI models come with very different licences. Some are as free as any open-source library; others cap the number of users, require attribution or forbid commercial use outright. If you ship a product or run a model for customers, the licence matters as much as the benchmark scores.
This article explains the common licence types in plain language. It is not legal advice: read the licence of the exact model you deploy, and ask your legal team when in doubt.
Open Weights Are Not the Same as Open Source
Most “open” models are open-weight: you can download and run the trained weights, but the training data and code often stay private, and the licence may restrict how you use them. In 2024 the Open Source Initiative published the Open Source AI Definition, which also expects enough information about the training data to rebuild the model. Only a few releases, such as NVIDIA’s Nemotron and AI2’s OLMo, publish data and recipes as well.
The Main Licence Types
Permissive: Apache 2.0 and MIT
These are standard open-source licences. You can use the model commercially, modify it, fine-tune it and redistribute it. Apache 2.0 also grants a patent licence and asks you to keep the licence and notice files. Families that use them for most models include Qwen, Gemma 4, gpt-oss, Granite, Phi (MIT), DeepSeek (MIT) and the small Mistral models.
Modified MIT
Some vendors take the MIT licence and add one condition, usually that very large products (by monthly users or revenue) must display the model’s name. Kimi K2.x, Mistral Medium 3.5 and Devstral 2 use variants of this. For most companies they behave like MIT, but read the added clause.
Community licences: Llama
The Llama licences allow commercial use, with conditions:
- Companies with more than 700 million monthly active users must ask Meta for a separate licence.
- You must show “Built with Llama” and include the licence when you distribute the model or a product based on it.
- An acceptable-use policy lists prohibited uses.
- Llama 3.x licences ask that fine-tuned models you distribute have names starting with “Llama”.
Meta’s newer Muse Glimmer model, by contrast, is Apache 2.0.
Vendor terms: Gemma Terms of Use and others
Gemma 1 to 3 use the Gemma Terms of Use. They allow commercial use but include a prohibited-use policy that you must pass on to your own users, and Google may update that policy. Gemma 4 moved to Apache 2.0. Other custom licences in this group include the Qwen Community License (Qwen3.8 Flash-Next), the Muse Model License, the Kimi K3 licences, the GLM-5.3 licence, the Mistral Ecosystem Open-Weight License (Mistral Large 4) and NVIDIA’s open model licences. Each is different, so read them one by one.
Non-commercial
Some weights are for research or evaluation only. Mistral’s Codestral 22B uses the Mistral Non-Production License, which does not allow use in a commercial product. Medical models such as MedGemma add their own terms.
Licences of Popular Model Families
This table is generated from the TecAdmin AI models directory and lists the licence of each family’s current models. Click a family for the licence of every individual model.
A Quick Checklist Before You Ship
- Find the licence of the exact model and version. Licences change between generations (Gemma 3 vs Gemma 4).
- Check whether a fine-tune or quantized download (for example a GGUF on Hugging Face) keeps the original licence. It must.
- Look for user or revenue thresholds, attribution rules and naming rules.
- Read the acceptable-use policy and pass it on to your users if the licence requires it.
- Keep a copy of the licence text with your deployment, as Apache 2.0 and most others require.
Not sure which model to pick? Compare candidates side by side, for example Qwen vs Llama or Gemma vs Llama, where the licence is part of the comparison.
How did it make you feel?
