Two open-weight models with similar benchmark scores can come with very different obligations. One licence asks you to pass on a copy of its text. Another asks you to bind your own users to a use policy. A third asks for permission once your product is large. This guide reads four published licences side by side so you can see those differences before you build on any of them. We read each text on 30 October 2026; licences change between versions, so treat the links as the authority.
The four licences at a glance
| Licence | Who publishes it | Grant | Conditions that travel with the model | Condition tied to scale |
|---|---|---|---|---|
| Apache License 2.0 | Apache Software Foundation | Perpetual, worldwide, royalty-free, irrevocable copyright licence[1] | Pass on the licence, mark changed files, keep notices, carry the NOTICE file if there is one[1] | None in the text |
| Gemma Terms of Use | Use under the agreement[2] | Pass on the agreement, mark changes, include the use restrictions as an enforceable provision[2] | None in the text we read | |
| Llama 4 Community License | Meta | Licence to the Llama materials | "Built with Llama" display, model names starting with "Llama", acceptable use policy[5] | Licence request above 700 million monthly active users[5] |
| Kimi K3 License | Moonshot AI | Permission grant worded like the MIT licence's opening[6] | Per the text, the grant is subject to its conditions[6] | Display "Kimi K3" in the product's interface above 100 million monthly active users or 20 million US dollars in monthly revenue[6] |
The table is a summary of clauses. Each licence has more text than a table can hold, and the notes below say where.
Apache License 2.0: the baseline
Apache 2.0 is a general-purpose software licence rather than an AI-specific one, which is why it is a useful baseline. Each contributor grants "a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable" copyright licence.[1] Section 4 lets you redistribute in any medium, with or without changes, if you give recipients a copy of the licence, mark modified files, keep the existing notices in the source form of derivative works and carry over the content of a NOTICE file when the work includes one.[1]
Two features matter in practice. First, the patent grant ends if you start patent litigation alleging that the work infringes: the text says patent licences granted to you "for that Work shall terminate" from the date the litigation is filed.[1] Only the patent licences end; the copyright licence is not mentioned in that clause. Second, section 6 grants no rights to the licensor's trade names or trademarks beyond reasonable use to describe the work's origin.[1]
As of our reading, the Qwen3.6-27B model card lists its licence as apache-2.0 in its metadata, without linking the full text on that page.[7] If you rely on it, open the licence file in the repository for the release you use.
Gemma Terms of Use: custom terms that bind your users
Google's Gemma Terms of Use (last modified 1 April 2026) work differently. Redistribution requires that recipients receive a copy of the agreement, that modified files carry change notices, and that you "include the use restrictions referenced in Section 3.2 as an enforceable provision" in your own agreement with downstream users.[2] Those restrictions point to the Gemma Prohibited Use Policy, whose categories include content that infringes others' rights, dangerous or illegal activity, misleading content including automated decisions in sensitive domains, and sexually explicit content with listed exceptions.[4] The policy opens with: "Google reserves the right to update this Gemma Prohibited Use Policy from time to time."[4]
Other clauses to note:
- Distribution includes hosting. The terms define distribution to include offering Gemma as a hosted service such as an API.[2]
- Outputs. The terms say outputs are not Model Derivatives and that Google claims no rights in them. You and your users are responsible for how outputs are used.[2]
- Termination. "Google may terminate this Agreement if you are in breach of any term of this Agreement." On termination you must stop using and distributing Gemma and its derivatives and delete your copies.[2]
The terms page says it covers the Gemma models listed in its appendix and that Gemma 4 is governed by a separate licence. The Gemma 4 licence page is headed "Apache License 2.0," last updated 1 April 2026.[3] Both pages are marked deprecated and point to gemma.ai, which we did not open, so confirm on the page that ships with the weights.
Llama 4 Community License: conditions at scale
The Llama 4 agreement has three conditions that the Apache text does not. If monthly active users exceeded 700 million in the preceding calendar month, you "must request a license from Meta."[5] Distributors must "prominently display 'Built with Llama' on a related website," and an AI model created or improved with the materials and then distributed must have a name beginning with "Llama."[5] Section 1.b also addresses using outputs of the materials to create, train, fine-tune or improve an AI model; read that sentence for the exact condition before you use outputs as training data.[5]
The endings differ from Apache 2.0. If you sue Meta or others claiming the materials infringe your intellectual property, "any licenses granted to you under this Agreement shall terminate" on the filing date, and Meta may terminate if you breach any term.[5] The agreement is governed by California law.[5] The acceptable use policy is incorporated by reference; we did not read its text this run, so we make no claims about what it prohibits.
Kimi K3 License: a permissive grant with a visibility rule
The Kimi K3 License opens with wording that resembles the MIT licence: permission "free of charge, to any person" obtaining the software, including the model weights and parameters, to use, copy, modify and create derivative works.[6] Its additional condition is a visibility rule rather than a permission gate: if the software or a derivative is used in a commercial product with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, "Kimi K3" must be prominently displayed in the product's user interface.[6] We covered the model's claims and release status in our Kimi K3.1 Pro article. Note that this licence belongs to the Kimi K3 repository; other Moonshot releases may differ.
What actually differs: an analysis
The following is our interpretation of the texts, not a legal opinion.
- What triggers the extra work. Apache 2.0 and Gemma impose duties on everyone who redistributes. Llama 4 and Kimi K3 impose their headline conditions only at very large scale, though Llama's naming and display conditions apply to ordinary distributors.
- What the condition asks for. Meta's threshold is a permission gate: you ask, and Meta decides. Moonshot's is a display duty: you show a name. The first can block a product; the second changes a screen.
- Who inherits the restrictions. Under Gemma's terms your customers are bound by use restrictions through you. Under Apache 2.0 they receive the licence and notices, not a use policy.
- What ends your rights. Apache 2.0 ends patent licences on a patent suit. Llama 4 ends all licences on an IP suit and permits termination for breach. Gemma's terms permit termination for breach and require deletion. A breach found late can mean retraining or migration.
- The policy can move. Gemma's prohibited use policy reserves Google's right to update it.[4] A product built on those terms depends on a document that can change.
A checklist before you ship
- Open the licence file shipped with the exact release, not the family's landing page, and record its version and date.
- List what your product does: hosts the model, ships it inside an app, fine-tunes it, uses outputs for training, serves other businesses.
- Match each activity to a clause. Flag any redistribution, naming, display, scale or output-use language.
- Check what your contracts must say. Gemma-style terms require you to pass restrictions to users.
- Estimate whether your user or revenue figures could ever reach a threshold, and note the date you would have to act.
- Read the incorporated documents, such as acceptable use policies, and note whether they can be updated unilaterally.
- Ask counsel about anything unclear. This guide does not replace that.
For the wider question of what "open" means, see open weights versus open source AI. If licence text feeds a vendor-risk review, our AI vendor data retention checklist shows the same read-the-source method for data terms.
Limits of this comparison
We read the texts but did not test enforcement, and we have no view on how courts treat them. We did not open the Llama acceptable use policy, the gemma.ai pages the deprecated pages point to, or the licences of other families. Secondary comparison sites disagree on licence details for several recent models, which is why we used only the licence texts and model card metadata here.
