Patent Risk in Llama, Qwen, DeepSeek, Kimi and Mistral

If your product runs on a model supplied by someone else, can you still face a patent claim?

Yes. A model or open source license may give you broad permission to use the code and weights, but it cannot clear patents owned by people who never granted rights under that license.

This is not a reason to avoid Llama, Qwen, DeepSeek, Kimi, Mistral, or open source AI. These five families show why the exact model license matters. Llama uses Meta’s custom terms, Qwen3 and Mistral 7B use Apache 2.0, DeepSeek R1 uses MIT, and Kimi K2 uses a modified MIT license.

The right starting point is to distinguish three questions that teams often collapse into one:

1. May we use and modify this code or model?
2. Which patent rights, if any, did its contributors grant us?
3. Could someone outside the project hold a patent relevant to what our product does?

The license usually answers the first question and sometimes part of the second. It cannot answer the third on its own.

Independent development is not always a defense

Copyright and patent law ask different questions. Copyright generally concerns whether protected expression was copied. A patent can be infringed by independently developed technology if the product or process falls within every element of a valid patent claim.

Under US patent law, direct infringement can occur when someone makes, uses, offers to sell, sells, or imports a patented invention without authority. Knowledge of the patent is generally not required for direct infringement. Induced and contributory infringement have additional requirements, however, and knowledge, notice, and intent can materially affect liability and remedies.

An AI company therefore cannot assume it is safe merely because its engineers never read the patent or because a mechanism arrived inside a freely available model or framework. The relevant questions are what the patent claims require, what the company actually does, and where the relevant acts occur.

Patent owners can also pursue commercial users instead of, or in addition to, the original software developer. Microsoft’s 2009 litigation against TomTom is a useful example. Microsoft asserted patents against the navigation device company over technology used in TomTom products, including file management functions related to Linux. The dispute ended in a patent agreement under which TomTom paid Microsoft for coverage.

That case does not prove that downstream enforcement is routine. It demonstrates that a commercial deployer can become the target.

AI increases the number of components that may need review: the base model, fine tuning framework, inference server, tokenizer, retrieval system, vector database, orchestration layer, and custom optimizations. Any of those layers may implement technology relevant to a patent claim.

What the common model licenses actually cover

Open source licenses primarily grant copyright permissions, although some also include express patent rights. No license can grant patents that the licensor does not own or control.

Model or license	Express patent protection	Important limitation
MIT and standard BSD	No express patent grant	Possible implied rights depend on the facts and should not be treated as general patent clearance
Apache 2.0	Yes, from each contributor	Covers only claims controlled by the contributor and necessarily infringed by that contribution, alone or combined with the work
Qwen3	Distributed under Apache 2.0	Receives Apache’s contributor grant, but not protection from unrelated third party patents
Mistral 7B v0.1	Distributed under Apache 2.0	Receives Apache’s contributor grant, but not protection from unrelated third party patents; other Mistral models may use different terms
DeepSeek R1	Distributed under MIT	Contains no express patent grant; distilled variants may also retain terms from their Qwen or Llama base models
Kimi K2.7 Code	Distributed under a modified MIT license	Contains no express patent grant and requires prominent Kimi K2 branding for commercial products above stated user or revenue thresholds
Llama 4	Limited rights under IP owned by Meta and embodied in the Llama materials	Does not license third party patents and remains subject to Meta’s conditions for use, redistribution, acceptable use, and large users

MIT and BSD contain no express patent grant

The MIT and standard BSD 3-Clause licenses give broad permission to use, copy, modify, and distribute software, but neither contains an express patent clause.

That is more precise than saying they grant nothing on patents. Arguments based on an implied license or estoppel may exist in particular circumstances, but they are not a substitute for an express grant or a freedom to operate review.

Apache 2.0 provides a useful but limited grant

Apache 2.0 expressly grants a patent license. Each contributor licenses patent claims that the contributor can license and that are necessarily infringed by its contribution alone or in combination with the work to which it was submitted.

That grant matters, but it leaves two obvious gaps. It does not cover patents owned by unrelated third parties. It also cannot provide a patent license where the contributor does not control the relevant patent.

Apache 2.0 includes defensive termination as well. If a licensee initiates patent litigation alleging that the work or an incorporated contribution infringes, the patent licenses granted to that licensee for the work terminate. That can discourage certain patent assertions, but it is not a general shield against every contributor, user, or external patent owner.

Qwen3 license and commercial use

The official Qwen3 repository states that all open weight Qwen3 models are licensed under Apache 2.0. Users therefore receive Apache’s express contributor patent grant, subject to the limits described above.

That distinction is commercially relevant because Qwen is no longer a niche family. The Hugging Face Spring 2026 open source AI report found more Qwen derivatives than models derived from Google and Meta families combined. Popularity does not expand the license, however. Qwen’s Apache terms still cannot license patents controlled by an unrelated owner.

Mistral licenses and commercial use

Mistral is not one uniform licensing category. Mistral 7B v0.1, for example, is distributed under Apache 2.0, so users receive the patent grant described above. That improves clarity for patents controlled by relevant contributors, but it does not establish freedom to operate against third party rights.

Always record the exact model and version used. Do not assume that the license attached to one Mistral model also governs another open weight or commercial Mistral offering.

DeepSeek R1 license and commercial use

The DeepSeek R1 repository distributes the repository and model weights under the MIT License. Unlike Apache 2.0, MIT contains no express patent grant. DeepSeek’s strong current use also makes it a practical example rather than a theoretical one. A 2026 study of open model adoption found DeepSeek leading its analyzed OpenRouter inference measurements.

The exact DeepSeek variant matters. The project states that distilled variants based on Qwen 2.5 retain the Apache 2.0 terms of their base models, while variants based on Llama retain the applicable Llama licenses. Those inherited terms affect what users may do with the model. They still do not provide a blanket license to patents owned by outsiders.

Kimi K2 license and commercial use

Moonshot AI releases Kimi K2.7 Code under its modified MIT license. It permits commercial use, modification, distribution, and sublicensing. It does not contain an express patent grant.

The modification matters for very large products. If the model or a derivative is used in a commercial product or service with more than 100 million monthly active users or more than $20 million in monthly revenue, the license requires the product to display “Kimi K2” prominently in its user interface. Teams should therefore describe it as modified MIT, not simply MIT.

Kimi is also commercially relevant rather than an edge case. Kimi K2.7 Code is available in GitHub Copilot Business and Enterprise, and its official Hugging Face release records substantial monthly downloads. Distribution through a major platform does not change the scope of Moonshot AI’s license or clear third party patents.

Llama 4 license and commercial use

Meta’s Llama 4 Community License grants a limited right under intellectual property and other rights owned by Meta and embodied in the Llama materials. It includes conditions on use and redistribution and a separate licensing requirement for organizations above the stated threshold of 700 million monthly active users.

The Meta grant is relevant to rights Meta controls. It cannot clear a patent owned by an unrelated company whose claims might read on a feature used in Llama or in the surrounding deployment stack. The Open Source Initiative also does not consider Meta’s Llama license open source.

The practical principle is simple: a license can convey only the rights its licensors are entitled to grant.

A published AI paper does not clear patent risk

Public disclosure of a research paper does not necessarily mean that every implementation described or inspired by it is free of patent rights.

Google holds active patent families directed to particular sequence transduction systems based on attention and associated with the development of transformer architectures. These include:

• US 10,452,978, which concerns neural networks for sequence transduction based on attention, and related European filings; and
• US 11,886,998, directed to particular decoder only sequence transduction systems based on attention.

This does not mean that Google owns every transformer or that every modern LLM infringes. A patent protects what its claims cover, not an architectural label. Infringement depends on claim construction and on whether the accused implementation contains every required element of at least one valid claim, either literally or, in some jurisdictions, through an equivalent.

Google also operates a patent pledge for specifically listed patents used in qualifying free or open source software. The two transformer patents identified above do not appear on the current list of pledged patents.

Separately, an Apache 2.0 patent grant may apply when qualifying Google contributions are actually used, but its scope must be analyzed from the license, the contributions, and the asserted claims. A history without enforcement should not be presented as either a license or a prediction of future conduct.

The wider AI stack also contains active patent applications and granted patents concerning areas such as model compression, quantization, retrieval, attention optimization, inference scheduling, and hardware acceleration. Their existence does not establish infringement. It does mean that open weights and patent clearance are different conclusions.

Five practical ways to reduce patent risk

1. Review freedom to operate against the actual product

A license audit tells you what your dependencies permit. It does not identify everything that third parties may be able to prohibit.

A freedom to operate review maps the product’s actual technical behavior against potentially relevant patent claims. For an AI product, that may include its training or fine tuning method, model architecture, inference optimizations, retrieval pipeline, caching, routing, and deployment configuration.

The review should prioritize granted patents that remain in force in commercially relevant countries while also tracking published pending applications that could create future risk. Jurisdiction is based on where legally relevant acts occur, not simply where the base model was trained. Development, hosting, use, offers, sales, and importation may take place in different countries, particularly in distributed cloud systems.

2. Maintain a software and model bill of materials

Track each model, framework, library, service, and material optimization used in production. Record the exact version, supplier, license, modifications, and model lineage.

A fine tuned, merged, or distilled model does not automatically inherit every patent risk associated with every ancestor. Patent exposure depends on what the resulting system actually implements. Lineage still matters because it helps investigators identify where relevant mechanisms came from and respond efficiently to a claim.

3. Prefer express patent grants, but understand their scope

An express patent grant such as Apache 2.0 is a meaningful factor when choosing between otherwise comparable components. It is not proof that the component has lower overall patent risk, and the size of a contributor is not a reliable proxy for freedom to operate.

Record who contributed the relevant functionality, which license version applies, and whether the patent grant covers modifications or combinations used in your deployment.

4. Negotiate indemnification carefully

Some commercial model and cloud providers offer intellectual property indemnities for particular services or enterprise customers. Coverage varies substantially.

For example, OpenAI’s current service terms extend specified IP indemnification to eligible API and enterprise output claims but contain important exclusions. Microsoft’s public Customer Copyright Commitment is principally framed around copyright claims. Neither should be reduced to a general promise that every patent affecting a customer’s complete AI stack is covered.

Indemnity does not make a patent disappear. It allocates some of the cost and responsibility under a contract, usually subject to procedures, exclusions, and sometimes liability caps. Review the exact agreement for patent coverage rather than assuming that a general IP indemnity protects the complete AI stack. Insurance and other negotiated methods of allocating risk may also be relevant.

5. Consider collective defensive networks

LOT Network addresses a specific risk. When a covered patent is transferred from a participating member to a qualifying patent assertion entity, LOT’s conditional license can protect other members from assertions based on that patent. It does not prevent ordinary patent disputes between operating companies.

Open Invention Network uses a different model. Participating organizations enter a cross licensing and nonaggression arrangement without royalties that covers the defined Linux System, alongside access to patents owned by OIN. Its protection is tied to that defined scope, not to every open source or AI technology.

Both networks can form part of a broader patent risk strategy, but neither replaces analysis of the specific product.

Do not overlook the option to design around a patent

When a review identifies a potentially blocking patent, litigation and licensing are not the only possible responses. Patents claim combinations of specific elements, not business goals in the abstract.

A claim chart can show whether the product contains every required element and identify technical features that might be changed. Replacing a claimed caching structure, routing method, or attention operation may avoid literal infringement, but the analysis must also consider other claims, related members of the patent family, and the doctrine of equivalents where applicable.

Designing around a patent is therefore an engineering and legal exercise. Done early, it may be considerably less expensive than changing a deployed product after receiving a claim letter.

The safer question to ask

The open source and open weight AI ecosystem is not free of patents. It combines generous access to code and models with patent rights that vary by contributor, owner, claim, implementation, and country.

The founders who manage this well do not avoid open models. They know exactly what is in the product, understand what each license actually grants, review the technical features with the highest risk, and keep options open for licensing or design changes.

The safer question is not “Is the model open source?”

It is: “Which rights do we have, which rights remain outside the license, and what does our deployed system actually do?”

Patenta supports semantic patent research and freedom to operate research across more than 160 million patents in 100+ jurisdictions, including FTO risk scoring and claim chart generation. Search the patent landscape around your AI product before a claim forces the question.

This article provides general information and is not legal advice. Patent scope, infringement, and licensing questions should be reviewed with qualified counsel for the relevant jurisdictions.