A fine-tuned model that ships inside your product carries three sets of terms with it: the base model’s license, the license of every dataset you tuned it on, and sometimes an acceptable use policy that the model’s maker can change after you have shipped. Most teams read none of them. This chapter reads them for you, clause by clause, and ends with the checklist I run before any open model goes into production.
This is chapter 3 of the Open Source AI Hub. It assumes the definitions from chapter 2. It is not legal advice: I am an architect, not a lawyer. Read the license text yourself, and for anything that ships to customers, ask counsel.
Two families: OSI-approved, and everything else
Every model license falls into one of two families: licenses approved by the Open Source Initiative, which grant the freedoms of the Open Source Definition without restriction on who may use the work or for what, and everything else. Apache-2.0 and MIT are in the first family. Every “community license”, “terms of use” and “responsible AI license” is in the second, because each restricts use by field, by scale or by policy, and the Open Source Definition’s sixth criterion forbids exactly that. That single test sorts most of the market before you read a word of the actual text.
Apache-2.0 and MIT for weights
MIT asks for one thing: keep the copyright and permission notice with the work. Apache-2.0 asks for the notice, a NOTICE file if the project ships one, a note on files you changed, and in return grants you an explicit patent license from every contributor, revoked only if you sue them over patents in the work. For model weights the patent grant matters more than it looks: training methods are patented, and Apache-2.0 is the only common permissive license that says the maker will not use those patents against you.
There is an open question underneath both licenses that nobody has settled: whether model weights are copyrightable at all. Weights are numbers produced by a process, not expression authored by a person, and courts have not ruled. The practical answer is that the license still binds you as a contract you accepted when you downloaded the files, so read it as one.
Models under these two licenses, as of September 2026: Qwen3 (Apache-2.0), Mistral’s open line including Mistral Small and Devstral (Apache-2.0), gpt-oss (Apache-2.0, with a short usage policy attached), IBM Granite (Apache-2.0), OLMo (Apache-2.0), Phi-4 (MIT), DeepSeek R1 and the V3 releases from March 2025 onward (MIT). For these, commercial use, modification and redistribution are all yours, with attribution.
The Llama Community License
The Llama Community License is Meta’s proprietary license for the Llama models, and its operative clauses are these. If the products or services of you or your affiliates had more than 700 million monthly active users in the calendar month before the model’s release, you must request a separate license from Meta, and you have no rights until Meta grants it. Use is subject to Meta’s Acceptable Use Policy, which is incorporated by reference and which Meta may update. If you redistribute the model or a derivative, you must include the license and display “Built with Llama”. Since Llama 3.1, any model you build from Llama weights must have “Llama” at the start of its name. And since Llama 3.2, the multimodal models grant no rights to individuals domiciled in, or companies with a principal place of business in, the European Union; Llama 4 carried a similar restriction.
Two changes over time are worth knowing. Llama 2 and Llama 3 prohibited using the model’s outputs to improve any other large language model; Llama 3.1 dropped that restriction in July 2024, which is why distilling from Llama became common afterwards. And the 700 million figure has been constant since Llama 2, so a handful of the largest companies in the world are excluded and everyone else is not.
My reading: for almost every company the terms are workable, and I deploy Llama models for clients. It is a generous proprietary license, not an open source one, because it discriminates by scale, by territory and by field of use. Say so in your documentation and nobody will be surprised later.
The Gemma Terms of Use
Google’s Gemma models ship under the Gemma Terms of Use, which incorporate a Prohibited Use Policy, require you to pass the same use restrictions on to anyone you distribute a derivative to, and reserve Google’s right to update the terms and to restrict use it considers to be in violation. The pass-through obligation is the clause that catches people: if you fine-tune Gemma and hand the result to a customer, the customer inherits Google’s policy, and you are responsible for telling them so. Check the current text before you rely on any of this; Google has revised the terms more than once.
Same conclusion as Llama: usable, often excellent, not open source.
RAIL and OpenRAIL
A Responsible AI License is a license that attaches a list of prohibited uses to an otherwise permissive grant and requires anyone who redistributes the model, or a derivative, to keep the same list attached. The BigScience OpenRAIL-M license on BLOOM (2022), the CreativeML OpenRAIL-M license on the early Stable Diffusion models (2022) and the BigCode OpenRAIL-M license on StarCoder and StarCoder2 are the ones you will meet. The prohibited uses are usually things nobody respectable would do, which is why the licenses feel harmless. The OSI rejects them anyway, on the same field-of-endeavor ground, and some researchers defend them as the only practical way to attach ethics to a release. Both positions are coherent; just do not call a RAIL-licensed model open source.
Mistral: three licenses from one company
Mistral is the clearest lesson that you must check the license per model, never per vendor. As of September 2026 its models fall into three tiers: Apache-2.0 for the open line (Mistral 7B, the Mixtral models, Mistral Small, Devstral and the small Magistral reasoning model); the Mistral Research License, which permits research and non-commercial use only, for models such as Mistral Large 2, Ministral 8B and Pixtral Large; and the Mistral AI Non-Production License for Codestral, which allows development and testing but not production use without a commercial agreement. The same “Mistral” name on the Hugging Face Hub can therefore mean three very different things. Read the license field on the specific repository, then read the file it points to.
Other vendor licenses in brief
NVIDIA Open Model License. Used for many Nemotron models: commercial use permitted, no claim on your outputs, with termination clauses including one for circumventing the model’s safety guardrails. Some Nemotron models are Llama derivatives and carry the Llama license instead.
Qwen. Most Qwen2.5 and all Qwen3 models are Apache-2.0, but the Qwen2.5 3B and 72B models used a separate Qwen license with a 100 million monthly active user clause. Check the size you download.
DeepSeek. The original DeepSeek V3 weights (December 2024) shipped under the DeepSeek Model License, with the code under MIT. DeepSeek R1 (January 2025) shipped under MIT, and the March 2025 update moved the V3 weights to MIT as well. A DeepSeek checkpoint from before March 2025 can therefore be under different terms from one after it.
Falcon. TII’s Falcon models use their own Apache-2.0-based license with an acceptable use policy attached.
Kimi K2. Moonshot’s Kimi K2 uses a modified MIT license that adds a prominent attribution requirement for products above a large user or revenue threshold; check the exact figures in the file.
The pattern across all of them: the closer a license gets to Apache-2.0 or MIT, the less there is to read; every clause a vendor adds is there to keep an option open for the vendor.
Datasets and provenance
Your fine-tune inherits the terms of the data as well as the terms of the base model. The Stanford Alpaca dataset is CC-BY-NC-4.0 because it was generated with an OpenAI model, so a model tuned on it cannot be used commercially even if its base was Apache-2.0. OpenAI’s and Anthropic’s terms of use both prohibit using their outputs to develop competing models, which affects any synthetic dataset you build with their APIs. On the other side, Ai2’s Dolma and Hugging Face’s FineWeb are released under the Open Data Commons Attribution license, and are safe to build on. Image models have their own history: the LAION-5B index is a set of links and captions under an attribution license, not a license to the images, and the Stanford Internet Observatory’s December 2023 report on illegal material in it is why several projects stopped using it.
The rule I apply: keep a provenance file for every model you ship, naming the base model and revision, every dataset and its license, and any distillation source. It takes ten minutes and it answers the first question every acquirer, auditor and enterprise customer will ask.
The EU AI Act, in three paragraphs
Regulation (EU) 2024/1689, the AI Act, entered into force on 1 August 2024, with the obligations for general-purpose AI models applying from 2 August 2025 and most remaining rules from 2 August 2026. It matters outside Europe because it applies to anyone placing a model or system on the EU market, which includes an Indian vendor selling to a European customer.
Article 53(2) exempts providers of general-purpose AI models released under a free and open source license from the technical documentation duties in the first two paragraphs of Article 53, but not from the duty to maintain a copyright policy or to publish a summary of training content. The exemption does not apply at all to models classed as having systemic risk, for which the presumption is a training run above ten to the power of twenty-five floating point operations. The recitals make clear that a model monetized through paid support, licensing or hosting does not count as free and open source for this purpose.
Whether a use-restricted license such as Llama’s qualifies as “free and open source” under the Act is debated, and the Commission’s guidance has continued to develop; treat any summary, including this one, as a snapshot as of September 2026 and read the current text and guidance before relying on the exemption.
How to check a model license before you ship
Eight steps. They take an hour the first time and ten minutes on every upgrade.
1. Identify the exact artifact
Write down the repository, the revision or commit, and whether you are using the original weights or a quantized fork by someone else. A quantized GGUF file on a community account is a derivative, and its license field is filled in by whoever uploaded it, not by the maker.
2. Read the license file, not the model card summary
The license field on the Hugging Face Hub is self-declared and sometimes wrong. Open the LICENSE file in the repository and, for vendor licenses, the vendor’s own canonical copy. Ten minutes of reading beats a year of assuming.
3. Check for use restrictions and an acceptable use policy
Search the text for “acceptable use”, “prohibited”, “restrict” and “policy”. If a policy is incorporated by reference, open it, save a dated copy, and note whether the vendor may change it after you have shipped.
4. Check scale thresholds and territory exclusions
Search for “monthly active users”, “revenue” and “European Union”. Decide whether the threshold could apply to you or to a customer you resell to, and whether a territory clause excludes any market you sell into.
5. Check redistribution, naming and attribution duties
If you will distribute the model or a derivative, note what must travel with it: the license text, a “Built with” notice, a name prefix, a NOTICE file. Put the obligations into your release checklist so they happen automatically.
6. Trace the lineage
List the base model, every dataset used to tune it, and any model whose outputs were distilled into it. The strictest term anywhere in that chain applies to the whole thing. This is where a permissive base model quietly becomes non-commercial.
7. Check the runtime and tooling licenses
The model is not the only artifact you ship. An observability tool under the Elastic License or a chat interface under the AGPL carries its own obligations for a hosted product. Chapter 5 lists the license of every tool I use.
8. Record it
Add a MODEL-LICENSES file to the repository with the date, the reviewer, the revision and the decision for every model and dataset, and re-run the eight steps on every upgrade. Licenses change between versions more often than people expect, as the DeepSeek and Llama histories above show.
Next steps
Continue to chapter 4, The Open Model Landscape, where the models named here are placed by size and task. The Private AI Architecture Checklist covers what to audit around the model once the license is settled, and if you would like the whole review done for you, that is part of private AI deployment consulting.