AI Models

Open-Weight Is Not Open-Source: 7 License Traps to Check Before You Build on a Free AI Model

Aditya Kumar JhaAditya Kumar JhaLinkedInAmazon·November 10, 2026·10 min read

Free to download is not free to use. Seven license traps in open-weight AI models, from user caps to naming rules, plus a safe-download checklist.

A model that you can download is not automatically a model that you can build a business on. "Open-weight" means the trained parameters are available. "Open source" has a specific meaning with more requirements, and between the two sits a patchwork of licenses with user caps, naming rules, acceptable-use policies and termination clauses. This guide is evergreen, but it matters this month because open-weight releases keep arriving, and the license is the part nobody reads. This is general information and not legal advice. Check the license text for the exact model and version you plan to use.

Insight

Quick summary: Open weights give you the numbers. Open source, as the Open Source Initiative defines it for AI, also requires information about the training data and the code used to train and run the system. Meta's Llama 4 license, for example, adds a 700 million monthly-user threshold, a "Built with Llama" requirement and a naming rule. Apache 2.0 is the permissive baseline, and Google moved Gemma to Apache 2.0 with Gemma 4 on April 2, 2026. Licenses vary within the same company's models, so read each model card. Also load weights in a safe format.

Trap 1: assuming open-weight means open source

The Open Source Initiative published its Open Source AI Definition 1.0 in October 2024. Its FAQ says the preferred form for modifying an AI system includes the software used to create the dataset, the training software, the training results (the parameters) and the legally shareable training data. It sorts data into four classes: open data must be shared, public and obtainable data must be disclosed with details on where to get it, and unshareable non-public data must be described in enough detail for others to build a dataset with the same structure. The definition requires "Data Information," not a full data dump. Critics say that is too lenient, while others want all data opened, and that debate is live. (Sources: OSI FAQ, TechCrunch on the definition)

In the FAQ, the OSI names Llama 2 as an example that fails, because it lacks required components or its legal terms conflict with open-source principles, and it adds that its assessments are not certifications. It does not address newer Llama versions, so do not assume the same verdict applies. The practical takeaway: when a headline says "open-source model," ask which of the four components is actually available. Many releases provide weights and a model card and little else. (Source: OSI FAQ)

Trap 2: the user-count cap

Meta's Llama 4 Community License says that if, on the Llama 4 version release date, the monthly active users of products or services made available by or for you or your affiliates exceeded 700 million in the preceding calendar month, you must request a license from Meta. Meta may grant or refuse it at its sole discretion, and you have no rights under the agreement until it does. For a startup this clause is invisible. For a large platform it is a gate controlled by a competitor. Earlier Llama licenses, such as Llama 3.1, contain a similar threshold. (Sources: Llama 4 license, Llama 3.1 license)

Trap 3: naming and attribution rules

The same Llama 4 license requires that if you distribute the materials or a product containing them, you prominently display "Built with Llama" on a related website, interface, blog post, about page or product documentation. If you use the materials or their outputs to create, train, fine-tune or improve a distributed AI model, the model's name must begin with "Llama." A "Notice" file with specified attribution text is also required. If you were planning a white-labeled product, these rules may collide with your branding. (Source: Llama 4 license)

Trap 4: the acceptable use policy hiding in a link

Under Llama 4's Section 1.b.iv, your use must comply with applicable laws, including trade compliance, and with Meta's Acceptable Use Policy, which is incorporated by reference. That means a separate document can change what you are allowed to do. Many open-weight releases include a use policy on the model card or in an appendix. Even permissive licenses can carry one. Trade coverage of DeepSeek's MIT-style license, for example, notes an attachment restricting illegal and hazardous uses, though that is a secondary-source description and you should read the text yourself. (Sources: Llama 4 license, dev.to license guide)

Trap 5: patent clauses that can end your license

Llama 4's Section 5(c) says that if you sue Meta or another entity claiming the materials or outputs infringe IP you own or can license, any licenses granted to you under the agreement terminate from the date the claim is filed. Apache 2.0 has a comparable defensive clause: each contributor grants a royalty-free patent license, and it ends for the Work if you file patent litigation alleging the Work infringes. Apache 2.0 also requires you to give recipients a copy of the license, mark modified files, keep notices and carry forward any NOTICE file. These are small obligations until you ship a product to a million users without them. (Sources: Llama 4 license, Apache 2.0)

Trap 6: assuming one company means one license

Google's earlier Gemma releases shipped under custom terms with a prohibited-use policy. Gemma 4, announced April 2, 2026, was released under Apache 2.0, and coverage described it as the first Gemma family under an OSI-approved license. Trade comparisons of Llama, Mistral, DeepSeek and Qwen licenses report that terms vary by release, with some Mistral models under Apache 2.0 and others under modified or non-production terms, and Qwen licenses differing by repository. Those comparisons conflict with each other on some models, which is the point: licenses change between versions, and fine-tuned derivatives need their own review. Always read the license file in the repository you actually download from. (Sources: Google Open Source Blog, The Next Web, SoftwareSeni comparison)

How common license types compare

License typeCommercial useTypical catchesExample
Apache 2.0YesKeep notices, mark changes, include the license. Patent grant ends if you sue over the Work.Gemma 4 (per Google and trade coverage)
MIT-styleYesVery short. Often described as lacking Apache's express patent grant. Some AI releases attach a use-restriction appendix.DeepSeek releases (per secondary reports)
Custom community licenseYes, with conditionsUser cap, attribution, naming rule, acceptable use policy, patent termination.Llama 4 community license
OSI Open Source AI DefinitionYesA definition, not a license. Requires data information, code and parameters under OSI-approved terms.Defined by the Open Source Initiative

Trap 7: the download itself

License risk is paperwork risk. Download risk is security risk, and the two are often confused. Hugging Face's own documentation warns that pickle, the default serialization format for PyTorch weights, can run arbitrary code when loaded, and advises loading models from users and organizations you trust. Its Hub scanner runs ClamAV and pickle import scans on uploaded files, and the docs state this is not foolproof and that users are responsible for checking safety. The same page points to safetensors as a safer weight format. Model hubs are also targets in their own right, as recent agent and vendor security incidents show. (Source: Hugging Face pickle scanning docs)

  • Prefer safetensors files over pickled checkpoints, and avoid loading files from accounts you cannot vouch for.
  • Download from the publisher's verified organization page and check the commit history.
  • Run unfamiliar models in a container or virtual machine with no access to credentials.
  • Pin the exact model revision and keep a copy of the license text you accepted that day.
  • Never trust a repository that asks you to run an install script before you can read the weights.

A 10-minute pre-build checklist

  • Find the license file in the repository, not the blog post. Note its name and version.
  • Search it for thresholds (users, revenue), attribution, naming and acceptable use references.
  • Check whether derivatives, fine-tunes or synthetic data made from outputs are restricted.
  • Confirm what the license says about patents and termination.
  • Check how any cap is measured (Llama 4's is tested as of the version release date), whether it applies to you, and whether a competitor controls that gate.
  • Record the model revision and the date you reviewed the terms.
Pro Tip

If a license is unclear, treat it as restrictive until a lawyer or the publisher says otherwise. Cheap clarification beats an expensive rebuild after you have shipped.

If you just want to try a range of AI models without sorting out licenses, hosting and downloads yourself, LumiChats gives you several AI models in one place. You can compare outputs first, then decide whether a self-hosted open-weight route is worth the legal and security homework.

Frequently Asked Questions
01What is the difference between open-weight and open-source AI?

Open-weight means the trained parameters are downloadable. The OSI's Open Source AI Definition also requires information about training data, plus the code used to train and run the system, under OSI-approved terms.

02Can I use Llama 4 commercially?

Generally yes under Meta's community license, with conditions: a 700 million monthly-user threshold, a Built with Llama notice, a naming rule for derived models and an acceptable use policy. Read the full text for your case.

03Is Apache 2.0 safe for commercial products?

It is the widely used permissive baseline. You must include the license, mark changes and keep notices, and its patent license ends if you sue over the Work.

04Did Google change the Gemma license?

Yes. Gemma 4, announced April 2, 2026, uses Apache 2.0. Earlier Gemma versions used Google's own terms with a prohibited-use policy.

05Are downloaded model files dangerous?

They can be. Hugging Face warns that pickle-format weights can execute code when loaded. Prefer safetensors, trusted publishers and isolated environments.

Was this article helpful?

Found this useful? Share it with someone who needs it.

Free to get started

Claude, GPT-5.4, Gemini —
all in one place.

Switch between 40+ AI models in a single conversation. No juggling tabs, no separate subscriptions. Pay only for what you use.

Start for free No credit card needed
Aditya Kumar Jha
Written by
Aditya Kumar JhaLinkedIn

Published author of six books and founder of LumiChats. Writes about AI tools, model comparisons, and how AI is reshaping work and education.

Keep reading

More guides for AI-powered students.