From Ritu Raj | Product & Market Analysis
Open Weight Licences Are Not Open Source, and 4 Clauses Decide If You Can Ship
On this page
Meta's Llama 4 licence stops granting you rights once your product passes 700 million monthly active users. That single sentence is why Llama fails the Open Source Definition. Most teams running an open source AI licence in production have never opened the file, and 4 clause types decide whether their deployment is actually compliant.
Key takeaways
- Open weights and open source are different legal states. The Open Source Initiative's definition requires freedom to use a system for any purpose. The Llama licence fails that test, and OSI has said so in writing.
- Four clause types do the real restricting. Scale triggers, field of use limits, flow-down duties and geographic carve-outs. Everything else in these files is warranty language and choice of law.
- The licence is versioned with the weights, not with the vendor. Llama 4 launched in April 2025 with a clause excluding companies based in the European Union from the multimodal rights. The text served at Meta's licence URL on 1 September 2026 no longer contains it.
- 2026 moved toward permissive terms, and the frontier moved back. Gemma 4 shipped under Apache 2.0 in April 2026, while Alibaba's largest release of the year carries a $50 million revenue trigger.
Open weights is a distribution choice. Open source is a licence.
Open weights means the parameters are published for download. You can run the model on hardware you control, fine tune it, and keep your data inside your own network. That is valuable, and it is not the same property as open source.
Open source is a state of the licence, not a state of the download button. The Open Source Initiative published its Open Source AI Definition 1.0 in October 2024. It requires four freedoms: to use the system for any purpose without asking permission, to study it, to modify it, and to share it. It also requires three components under approved terms: data information detailed enough to rebuild an equivalent system, the training and inference code, and the parameters.
Measured against that, very little of what is marketed as open source AI qualifies. Llama fails at the first freedom, long before anyone reaches the training data question. OSI wrote in February 2025 that the Llama licence fails freedom 0, discriminates against users, and restricts fields of endeavour.
The definition argument is the least useful part of this story. Whether a licence earns the words "open source" is a question for standards bodies. Whether it lets you ship is a question for you, and the answer changes clause by clause.
The reason to care now is volume. Open weight models are no longer a research preference. They are carrying real production traffic. Whether they are ready for it is a separate question, covered in the production comparison of open weight models.
The 4 clauses that decide whether you can ship
Read enough of these files and the same four shapes appear. The rest is warranty disclaimer and governing law.
1. Scale triggers
Llama 4, Section 2, is the best known example. The trigger is measured on the version release date, for your products or services in the preceding calendar month. Above 700 million monthly active users, "you must request a license from Meta, which Meta may grant to you in its sole discretion." The operative phrase is sole discretion, not the number.
Chinese labs have adopted the same shape with different metrics and much lighter consequences. Moonshot AI's Kimi K2 ships under a modified MIT licence. Above 100 million monthly active users or $20 million in monthly revenue, you must display "Kimi K2" on the user interface of that product. That is a branding duty, not a permission gate.
Alibaba's largest release of 2026 goes the other way. Reporting on the Qwen3.8-Max licence describes a separate commercial licence requirement for model-as-a-service and AI assistant products above $50 million of revenue in any consecutive 12 month period. The smaller Qwen models stay on Apache 2.0. The wider trade-offs of building on Chinese open weights are set out in the analysis of sovereignty cost in Chinese open models.
Three vendors, three trigger metrics, three consequences. A policy that says "we only use permissively licensed models" does not survive contact with that variety.
2. Field of use restrictions
Almost every non-Apache licence in this category incorporates an acceptable use policy by reference. Gemma's terms bind you to the Gemma Prohibited Use Policy. Llama binds you to the Llama Acceptable Use Policy. You accept a document that is not in the file you read.
Two properties of these policies matter more than their contents. They live at a URL the vendor controls and can be revised without your involvement. They also reach model derivatives, which for Gemma includes models trained on synthetic data the model produced.
Google's terms go further than any software licence most engineering teams have signed. Section 3.2 states that "Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement." The standard for acting is the vendor's own reasonable belief. How this compares against the hosted model APIs is covered in the comparison of model provider terms.
3. Flow-down and branding duties
This is the clause that catches product companies, and in my experience it is the one missed most often. It creates work outside engineering entirely.
Gemma's Section 3.1 requires four things of anyone distributing Gemma or a derivative. Include the use restrictions "as an enforceable provision in any agreement" governing downstream use. Give every recipient a copy of the agreement. Mark modified files with prominent notices. Ship a notice file pointing at the terms.
Read the first item again. If you embed a Gemma derivative in software you sell, your own customer contract has to carry Google's restrictions as an enforceable term. That is an edit to your commercial paper, not a line in a NOTICE file, and it needs a lawyer rather than a build step.
Llama's version is lighter and more visible. You must display "Built with Llama" prominently on a website, interface, blog post or documentation. Any model you distribute that was created, trained or improved using Llama materials must carry "Llama" at the beginning of its name. Your model naming is therefore partly decided by your supplier.
4. Geographic carve-outs
When Llama 4 shipped in April 2025, the licence withheld the multimodal rights from individuals domiciled in the European Union and from companies with a principal place of business there. End users of a product built outside the EU were exempt. The clause restricted builders, not users, and every Llama 4 model was multimodal.
Geography is the clause type least likely to appear in a compliance checklist, because software licences historically did not carry it. That assumption no longer holds.
What the major licences actually say
The table below is the version of this post worth keeping. It compares licence families rather than individual models, because models move between families and families are stable.
| Licence | Scale trigger | Use restrictions | Downstream duties | Open source by OSI test |
|---|---|---|---|---|
| Apache 2.0 Gemma 4, Qwen 3, Mistral Large 3, gpt-oss weights | None | None in the licence | Preserve notices, state modifications | Licence yes, full system no |
| MIT DeepSeek weights | None | None | Preserve the copyright notice | Licence yes, full system no |
| Modified MIT Kimi K2 | 100M users or $20M monthly revenue | None added | Display "Kimi K2" above the threshold | No |
| Llama Community Licence | 700M monthly active users, Meta's sole discretion above it | Acceptable Use Policy, incorporated by reference | "Built with Llama", "Llama" prefix on derivative model names | No |
| Gemma Terms of Use Gemma 3 and earlier | None | Prohibited Use Policy, plus a remote restriction right | Restrictions must be an enforceable provision downstream | No |
| Research and non-production licences | Not applicable, commercial use excluded | Research or evaluation only | Redistribution generally limited | No |
The last column is deliberately split for the permissive rows. Apache 2.0 and MIT are approved open source licences, but a model shipped under them still fails the Open Source AI Definition unless the training data information and code are also released. Almost none are.
The permissive column is genuinely permissive. Apache 2.0 carries no usage cap, no royalty, no field of use limit, and it includes an explicit patent grant that MIT does not. If your legal position needs to be simple, that patent grant is worth choosing for on its own.
Moving from an Apache 2.0 model to a Llama derivative is not a swap of one dependency for another. It adds an acceptable use policy you do not control, a naming rule for your own product, and a permission gate that some future version of your company might hit.
The licence is versioned with the weights, not with the vendor
Meta publishes a new agreement with every model family. Llama 2, Llama 3, Llama 3.1, Llama 4 and Llama 5 each have their own licence at their own URL. The terms are not identical between them, and the differences do not appear in release notes.
The European Union clause shows how far this goes. The Llama 4 exclusion was reported widely at launch in April 2025. I fetched the text served at Meta's Llama 4 licence URL on 1 September 2026 and searched it for "European Union" and for "domiciled". Neither string was present.
I am not suggesting Meta acted improperly. Removing a restriction is a change in the licensee's favour, and the regulatory context that produced it has moved on. The point is narrower and holds in either direction: the document that governs your deployment sits at a URL somebody else controls, and it changed without telling you.
Llama 5 arrived on 8 April 2026 under an updated community licence that commentary describes as friendlier to medium and large companies than the Llama 4 terms. I could not retrieve the licence text from Meta's canonical licence paths while writing this, which is the practical case for the habit below.
What to archive, and when
Archive the licence file, the referenced use policy and the model card at the moment you take the weights, with a timestamp and a hash. Store them beside the model artefact, not in a wiki page that will be reorganised.
When somebody asks in 2028 which terms you accepted, the answer should be a file you kept rather than a URL you hope still resolves. This is ordinary practice for third party code and not yet for weights. That gap is the cheapest thing on this page to close.
The licence now decides a regulatory question too
Article 53(2) of the EU AI Act exempts general purpose AI models released under a free and open-source licence from parts of the provider obligations. The exemption has three conditions and all three must hold. The licence must allow access, use, modification and distribution. The parameters, architecture information and usage information must be public. The model must not be provided against a price or otherwise monetised.
Apache 2.0, MIT and OpenMDW clear the first condition. A licence carrying a scale trigger, a research-only limit or a field of use restriction does not clear it.
Two obligations survive the exemption regardless. Article 53(1)(c) requires a policy for complying with EU copyright law. Article 53(1)(d) requires a public summary of training content using the official template. New models have been in scope since 2 August 2025, and models already on the market have until 2 August 2027.
The consequence is concrete. Fine tune a restricted-licence model, place it on the EU market as a provider, and your licence choice has moved you out of an exemption you could have had for free. That is a compliance cost created by a document nobody in your company negotiated, and it stacks on top of the disclosure work described in the EU AI Act transparency checklist.
What a breach actually costs you
Nothing dramatic happens on day one, which is exactly the problem. There is no invoice and no warning email.
Llama's Section 6 reads: "Meta may terminate this Agreement if you are in breach of any term or condition of this Agreement. Upon termination of this Agreement, you shall delete and cease use of the Llama Materials." The remedy is not a fee. It is deletion.
For a product with a fine tuned model in the serving path, that is a re-platform under time pressure. The base weights go, the derivatives built on them go, and your customer commitments stay exactly where they were.
The surrounding sections are equally worth reading before signing. Section 3 disclaims all warranties. Section 4 excludes lost profits and consequential damages. Section 7 sets California law with exclusive jurisdiction in California courts. For an Indian or European company, that last one converts a small dispute into an expensive one.
The indemnity you do not have
A commercial model API contract usually carries some form of intellectual property indemnity. Open weight licences carry none. You take the weights as they are, including any exposure attached to how they were trained. That exposure is an open question in several live cases, discussed in the review of AI copyright cases over training data.
That does not make open weights the wrong choice. It makes the comparison different from the one most build-versus-buy spreadsheets run, which usually stops at cost per token and latency. Add a row for indemnity and a row for termination remedy, and the ranking sometimes changes. The contract clauses worth insisting on either way are set out in the CFO guide to AI contract clauses.
Where this argument is weakest
Two objections to this post are correct.
The scale triggers catch almost nobody
700 million monthly active users is a threshold roughly a dozen companies on earth reach. $50 million of revenue in a rolling year is a real business, but it is a large one. For most readers, the scale trigger is theatre. If your objection is that this post over-weights a clause you will never hit, the arithmetic is on your side.
The clauses that reach an ordinary company are the flow-down duties and the incorporated use policies, and those get the least attention. I still think the scale triggers matter for one indirect reason: they are what makes a licence non-open, which removes the EU AI Act exemption, which is a cost you pay at any size.
The direction of travel is genuinely good
2026 has been the most permissive year for open weight licensing so far. Gemma 4 shipped on 2 April 2026 under Apache 2.0, replacing the custom Gemma terms quoted above. Qwen 3 moved to Apache 2.0. DeepSeek moved to MIT. Mistral Large 3 shipped Apache 2.0 in December 2025. The gpt-oss weights are Apache 2.0 with a short usage policy attached.
A post that reads the restrictive licences and ignores that trend would be misleading. The honest summary is that the mid-range has become permissive while several frontier releases have not, so the licence question follows you upward as you buy more capability.
One more limitation, on this post's own evidence. Counts circulating in 2026 of how many "open" models are genuinely permissive come from single-author audits with small samples, so this post does not cite one. Every clause quoted above was read in the licence text itself on 1 September 2026, which is a narrower claim and a checkable one.
A 30 minute licence audit you can run this week
This is a text search exercise, not a legal project. Run it against every model in your serving path, then hand the exceptions to counsel.
| Check | Where to look | A bad answer sounds like |
|---|---|---|
| The exact licence name and version | The LICENSE file shipped with the weights, not the model card summary | "It is on Hugging Face, so it is open" |
| Scale triggers | Search the text for "monthly active", "revenue", "affiliates" | Nobody has opened the file since the proof of concept |
| Incorporated policies | Search for "acceptable use", "prohibited use", "by reference" | The policy URL was never archived |
| Flow-down duties | Search for "distribute", "enforceable provision", "notice" | Customer contracts say nothing about model terms |
| Derivative scope | Whether outputs and synthetic data count as derivatives | Fine tunes are assumed unencumbered |
| Version pinning | Which licence version you accepted, and on what date | The checkpoint was upgraded and the old assessment reused |
| EU AI Act position | Whether the licence meets all three Article 53(2) conditions | Assumed exempt because the weights are downloadable |
The output is a short table per model, and it belongs with your model inventory rather than in a legal folder. If you already maintain a bill of materials for AI components, add licence name, licence version, archived policy URL and threshold status as four columns. The structure is in the AI bill of materials disclosure template.
Two of these checks find nearly everything. Derivative scope decides whether your fine tune inherited the terms. Version pinning decides whether the assessment you did last year still describes what you are running today.
Frequently asked questions
Is Llama open source?
No. The Llama Community License is not approved by the Open Source Initiative, which stated in February 2025 that the licence fails freedom 0, discriminates against users and restricts fields of endeavour. Llama is open weights, meaning the parameters are published for download and can be run and fine tuned on your own hardware. The two terms describe different things and only one of them is a licensing status.
Can I use Llama commercially for free?
Yes, for almost every company. The Llama 4 licence grants royalty-free commercial use below 700 million monthly active users, measured on the version release date for the preceding calendar month and including affiliates. Above that, you must request a licence from Meta, which Meta may grant at its sole discretion. You must also display "Built with Llama" and follow the Acceptable Use Policy incorporated into the agreement.
What is the difference between open weights and open source?
Open weights describes distribution: the model parameters are published and you can download and run them. Open source describes the licence: it must permit use for any purpose, study, modification and sharing. The Open Source AI Definition also expects training data information and code to be released. Most open weight models are published under licences that fail at least one of those conditions.
Which open weight models have no commercial restrictions?
Models shipped under unmodified Apache 2.0 or MIT carry no usage cap, royalty or field of use limit. That currently includes Gemma 4, Qwen 3, Mistral Large 3, the gpt-oss weights and the DeepSeek weights. Apache 2.0 additionally grants patent rights that MIT does not. Always confirm against the LICENSE file shipped with the specific checkpoint, because vendors change licences between model versions.
Do open source AI licences qualify for the EU AI Act exemption?
Only some do. Article 53(2) requires a licence permitting access, use, modification and distribution, public parameters and architecture information, and no monetisation of the model. Apache 2.0 and MIT satisfy the licence condition. Scale triggers, research-only terms and field of use restrictions do not. The copyright policy and the public training data summary obligations apply either way.
What happens if I breach an open weight model licence?
Under the Llama agreement, Meta may terminate the licence for breach of any term, and on termination you must delete the Llama materials and stop using them. There is no fee that cures the problem. For a deployed product that means removing the base model and anything derived from it, while your customer obligations continue unchanged. The agreement also disclaims warranties and sets California jurisdiction.
Where to start
Open a terminal before you open a meeting invite. Run the seven checks above against your two highest traffic models, and archive the licence file and the referenced use policy next to each checkpoint with today's date on them. That is under an hour, and it converts a vague exposure into a two-row table.
Then ask one question of whoever owns your customer paper: does any agreement we sell under carry an obligation we inherited from a model licence? If the answer is a pause, you have found the gap that the flow-down clauses create, and it is cheaper to close before the next enterprise deal than during its legal review.
Related work
Licence terms are one input to the open weights decision. The others are covered in the production readiness comparison and in the review of hosted model provider terms.
References
- Meta, Llama 4 Community License Agreement, version effective 5 April 2025. Used for the 700 million user threshold, attribution and naming duties, termination, warranty and governing law clauses. Text read 1 September 2026.
- Google, Gemma Terms of Use, last updated 1 April 2026, and Google Open Source Blog, Gemma 4: Expanding the Gemmaverse with Apache 2.0, 2026. Used for Sections 3.1 and 3.2, the flow-down duties, the remote restriction right and the Gemma 4 licence change.
- Moonshot AI, Kimi K2 LICENSE, modified MIT. Used for the 100 million user and $20 million monthly revenue display trigger.
- Open Source Initiative, The Open Source AI Definition 1.0, October 2024, and Meta's LLaMa license is still not Open Source, Jordan Maris, 18 February 2025. Used for the four freedoms, the data, code and parameters requirement, and the freedom 0 and Open Source Definition points 5 and 6 objections.
- Hugging Face, What Open-Source Developers Need to Know about the EU AI Act's Rules for GPAI Models. Used for the Article 53(2) conditions, the surviving obligations and the compliance dates.
- Vercel, AI Gateway Production Index, 13 July 2026. Used for the 29% token share and the sub-4% spend share for June 2026.
- ioplus, European Union excluded from Llama 4 multimodal models, 8 April 2025. Used for the original EU carve-out and the end user exception.
- Forkast, Alibaba's Qwen 3.8 licence is a platform play, 13 August 2026. Used for the $50 million revenue trigger.
The weakest part of this source base is the Qwen3.8-Max threshold, which rests on secondary reporting because the licence text was not retrievable from a vendor-hosted repository on 1 September 2026. Every other clause quoted here was read in the primary licence document on that date. Licence text can change at a stable URL without notice, which is the argument of the piece and also a caveat on it.
Related reading