Insights

Open-Source LLM Licenses: What Commercial Use Allows

Most open-weight models allow commercial use, but the conditions differ. A permissive license asks for a notice; a community license can add user caps, naming rules and a use policy.

Share:
Open-Source LLM Licenses: What Commercial Use Allows

Almost every popular open-weight model can be used by a business, but "open" covers two very different kinds of open source LLM license, and commercial use means different things under each. Under Apache 2.0 or MIT, the main duty is to keep a copyright notice with anything you redistribute. Under a custom community license, such as the ones Meta publishes for Llama, you also accept a user threshold, naming and attribution rules, and an acceptable use policy that is part of the contract. What follows is general information, not legal advice; a lawyer should review any license your product depends on.

Three kinds of license you will meet

Model licenses fall into three practical groups. The table summarizes them as of October 2026; check the current license file of the exact model before relying on it.

Kind Examples (October 2026) Commercial use What it adds
Permissive Apache 2.0 (Qwen3, gpt-oss, Gemma 4), MIT (DeepSeek-R1) Yes Keep notices; Apache adds a patent clause
Custom community Llama 3.1 and Llama 4 Community Licenses, earlier Gemma Terms of Use Yes, with conditions Use policy, attribution, sometimes a size threshold
Restricted Research or non-production licenses some vendors use for certain models No, or only under a separate paid license Commercial rights sold separately

A vendor can sit in more than one row. Mistral's help center says "most" of its open models use Apache 2.0, while "certain models" use a modified MIT license that requires companies with monthly revenue above $20 million to obtain a commercial license or use the model through Mistral's own platform. That is $240 million a year (20 x 12), far above most readers of this article, but the point stands: the vendor name tells you nothing. The license file in the model's repository does.

A note on the word itself. The Open Source Initiative's Open Source AI Definition 1.0 requires freedom to "use the system for any purpose and without having to ask for permission," plus detailed information about the training data. Most open-weight releases do not publish that data information, and custom licenses with use restrictions fail the "any purpose" test. "Open-weight" is the more accurate term for most of what businesses download.

What a permissive license asks of you

Apache 2.0 grants a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable" license to reproduce, modify, sublicense and distribute the work. It does not ask what you use the model for. Its conditions bite only when you redistribute: give recipients a copy of the license, mark files you changed, keep existing copyright and attribution notices, and include the contents of any NOTICE file.

Apache 2.0 also includes a patent grant with a termination clause. If you sue anyone claiming the work infringes a patent, "any patent licenses granted to You under this License for that Work shall terminate." For most businesses this never matters, but it is the main legal difference from MIT.

MIT is shorter still. Its one condition is that "the above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software." DeepSeek's R1 model card goes further than the bare license and states that the series supports commercial use, modification and derivative works, "including, but not limited to, distillation for training other LLMs."

For a company that runs a permissive model on its own server for its own staff, nothing is being redistributed, so in practice the conditions are light. They come into play when you ship the weights, or a fine-tuned version, to someone else.

What a community license adds

Meta's Llama licenses are the best-known custom licenses and a good template for what to look for. Four clauses do most of the work.

A size threshold

Section 2 of both the Llama 3.1 and Llama 4 licenses says that if your products or services had "greater than 700 million monthly active users in the preceding calendar month" on the version release date, you must request a license from Meta, which it may grant or refuse. No company of 5 to 500 people comes near this, but read thresholds carefully in other licenses, since they may count revenue, users or employees, and may be far lower.

Attribution and naming

Section 1.b.i applies "if you distribute or make available the Llama Materials (or any derivative works thereof), or a product or service" that contains them. You must provide a copy of the agreement and "prominently display 'Built with Llama'" on a related website or in the product. If you fine-tune a Llama model and distribute or make it available, both licenses say you "shall also include 'Llama' at the beginning of any such AI model name." There is also a required notice line naming the license and Meta's copyright, for example "Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved."

The phrase "make available" matters. A customer-facing chatbot built on Llama is arguably a product or service that makes the materials available, even if the weights never leave your server. Purely internal staff use is a weaker case, but the license does not draw that line for you.

An acceptable use policy

Section 1.b.iv requires compliance with applicable law and with Meta's Acceptable Use Policy, which is incorporated by reference. The Llama 4 policy groups prohibited uses into broad areas such as illegal activity, uses that endanger people, deception, and failing to disclose known risks of an AI system to end users. Because the policy sits outside the license file and can be updated, record the version you accepted.

Geographic carve-outs

The Llama 4 use policy contains a clause many readers miss. For "any multimodal models included in Llama 4," the license rights "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." The restriction does not apply to end users of a product that incorporates the model. A European company building on a Llama 4 vision model has a problem that a US company does not.

Use restrictions that travel downstream

Custom licenses often require you to pass their restrictions on. The Gemma Terms of Use, which still govern Gemma models before Gemma 4, say a distributor must "include the use restrictions referenced in Section 3.2 as an enforceable provision in any agreement" governing the model or its derivatives. Section 3.2 refers to the Gemma Prohibited Use Policy and adds that Google "reserves the right to restrict (remotely or otherwise) usage" it believes violates the agreement. The same terms state that Google "claims no rights in Outputs" you generate.

Google moved Gemma 4 to Apache 2.0 in April 2026, describing it as the first Gemma release under an OSI-approved license. A team that upgrades from Gemma 3 to Gemma 4 therefore changes license type, which is a good reason to treat every model version change as a license review, alongside the testing described in the guide to updating on-premise AI models.

The license on the file may not be the one you think

Three traps catch teams that read only the model's name.

  • Derivatives inherit the base model's license. DeepSeek-R1 itself is MIT, but its model card states that the distilled versions carry their base models' licenses: the Qwen-based distills come from Apache 2.0 models, while the Llama-based ones fall under Meta's Llama licenses. A model called "DeepSeek" can carry Llama obligations.
  • Re-uploads drop paperwork. A quantized copy posted by a third party may omit the license file, the NOTICE file or the use policy link. The obligations still apply. Download from the original publisher, or confirm the base model and its license before you use a converted copy; the article on model formats and verifying downloads covers checking that a file is what the publisher released.
  • Code and weights can differ. Some repositories license the inference code under one license and the weights under another. Read the statement that covers the weights.

A five-question reading method

Before a model goes into anything beyond experiments, open its license and answer these in writing:

  1. Is commercial use allowed, and on what terms? Look for words like "non-commercial", "research", "non-production" and any revenue or user threshold.
  2. What triggers the redistribution duties? Shipping weights, making a service available, or both. Map that to how you plan to deploy.
  3. What attribution is required? Notice files, "Built with" labels, naming rules for fine-tuned models.
  4. Is there a use policy, and where does it live? Note its URL and the date you read it, since an external policy can change.
  5. What about derivatives and outputs? Whether fine-tunes, distillation and generated text carry conditions, and whether the license says anything about using outputs to train other models.

Keep the answers, with the license version and download date, in the same place you record which model is deployed. License fit is one of the first filters in the method for picking an open-source LLM for business; a restriction found after the pilot is far more expensive than one found before the shortlist. Vision models carry their own licensing issues, covered separately in the article on YOLO licensing for commercial use.

Frequently asked questions

Can one model's outputs be used to train another model?

It depends on the license. DeepSeek's model card names distillation as a permitted use of R1. The Llama 3.1 and Llama 4 licenses allow it but attach a condition: if you use Llama outputs to "create, train, fine tune, or otherwise improve an AI model" that is distributed or made available, its name must begin with "Llama". A model trained on Llama outputs therefore carries a Llama obligation even if its weights came from elsewhere.

Does a business own the text a licensed model generates?

Several licenses address this directly; the earlier Gemma terms say Google claims no rights in outputs. Whether anyone holds copyright in machine-generated text is a separate legal question that depends on jurisdiction. The license only settles whether the model's publisher asserts a claim.

Can a model license change after download?

The license that came with the version you downloaded generally stays with that copy, which is why recording the version and date matters. External documents such as acceptable use policies are often incorporated by reference and can be updated, so check them when you upgrade or expand use.

Is a model with a custom license safe to use commercially?

Often yes, provided you meet its conditions: thresholds, attribution, and the use policy. The risk is not the license type but unread clauses, such as the Llama 4 restriction on multimodal models for companies based in the EU. Have counsel review any license a revenue-bearing product depends on.

Ready to Modernize Your Business?

Get your AI automation roadmap in minutes, not months.

Analyze Your Workflows →