Gemma Terms of Use
The Gemma Terms of Use are the custom license agreement under which Google has distributed the weights of its Gemma models since the family's first release on February 21, 2024, covering every generation from Gemma 1 through Gemma 3n and a list of derivative models such as PaliGemma, CodeGemma and ShieldGemma.[1][9] They incorporate by reference a separate document, the Gemma Prohibited Use Policy, which lists the uses that are not allowed.[1][2] The agreement lets anyone use, reproduce, modify and redistribute the models and their derivatives, including for commercial purposes and with no user or revenue threshold. In return, anyone who redistributes a model must pass Google's use restrictions on as an enforceable term, and Google reserves the right to "restrict (remotely or otherwise) usage" that it reasonably believes violates the agreement.[1]
Because the terms limit how the models may be used, they are not an open source license in the sense used by the Open Source Initiative, and Google described Gemma at launch as an "open model" rather than as open source.[10][17] The custom terms drew criticism from developers, researchers and the Open Source Initiative's head of community, who said they deterred commercial adoption.[11] Google announced Gemma 4 on April 2, 2026 under the Apache License 2.0 instead (its release log dates the weights to March 31).[32] A revision of the Gemma Terms of Use dated April 1, 2026 says that they apply only to the models listed in their appendix and points Gemma 4 users to the Apache license.[1][8][21] As of November 2026 the Gemma Terms of Use still govern the earlier models. Google's medical models MedGemma and TxGemma are distributed under a separate agreement modeled on them, the Health AI Developer Foundations Terms of Use.[14][16]
Background
Google announced Gemma on February 21, 2024 as "a family of lightweight, state-of-the-art open models" built from the research used for its Gemini models. The launch post said that the "Terms of use permit responsible commercial usage and distribution for all organizations, regardless of size."[9] Google offered free access through Kaggle and integrations with tools such as Hugging Face Transformers.[9]
On the same day, Google's Open Source Programs Office and Google DeepMind's responsibility team explained the choice of terms in a post on the Google Open Source Blog. They wrote that Gemma was being released as what the industry had begun to call "open models", and that "Open models feature free access to the model weights, but terms of use, redistribution, and variant ownership vary according to a model's specific terms of use, which may not be based on an open-source license." The post said the terms made the models freely available to individual developers, researchers and commercial users for access and redistribution, that users were free to create and publish variants, and that Google was "precise about the language" it used to describe Gemma because it was also a supporter of open source. It argued that "existing open-source concepts can't always be directly applied to AI systems".[10]
Revision history
Google has published four dated versions of the Gemma Terms of Use. Its own archive page lists the two 2024 versions as previous versions; a March 2025 version that preceded the current text survives in the Internet Archive's Wayback Machine.[3][6]
| Version ("Last modified") | Main change | Sources |
|---|---|---|
| February 21, 2024 | First text. "Gemma" meant the models "identified at ai.google.dev/gemma". Section 4.1 required licensees to make "reasonable efforts to use the latest version of Gemma". | [4] |
| April 1, 2024 | Section 4.1 cut to one sentence, "Google may update Gemma from time to time.", dropping the duty to use the latest version | [3][5] |
| March 24, 2025 | "Gemma" redefined as the models "identified in the Appendix"; an appendix listing 16 models added | [6] |
| April 1, 2026 | New header: the terms apply to the models in the appendix, and Gemma 4 users are pointed to the Gemma 4 license | [1][8] |
The appendix has grown without a change to the "Last modified" date. The first appendix listed Gemma 1, Gemma 1.1, Gemma 2, Gemma 3 and twelve other entries. Archived copies show Gemma 3n added by May 22, 2025, T5Gemma by July 30, 2025, EmbeddingGemma by September 7, 2025, VaultGemma by September 13, 2025, FunctionGemma and T5Gemma 2 by December 23, 2025, and TranslateGemma by January 20, 2026.[6][7] FunctionGemma has been listed twice since it was added.[1][7]
The Prohibited Use Policy has kept its original date, February 21, 2024, and its wording in November 2026 matches the copy archived on launch day, although the policy says Google may update it "from time to time".[2][4]
By November 2026 Google had moved the Gemma documentation to the gemma.ai domain: ai.google.dev/gemma/terms redirects to gemma.ai/resources/terms. The current text gives gemma.ai addresses for the notice and the Prohibited Use Policy, while copies archived as late as November 6, 2026 gave the ai.google.dev addresses under the same April 1, 2026 date.[1][8] The policy address written in the current text, gemma.ai/resources/prohibited_use_policy, does not load; the policy is published at gemma.ai/resources/prohibited-use.[1][2]
What the terms grant
The agreement is accepted by use. Its preamble says that anyone "using, reproducing, modifying, distributing, performing or displaying any portion or element of Gemma" or of a Model Derivative, including through a hosted service, agrees to be bound. "Gemma" is defined as the listed models "regardless of the source that you obtained it from", so the terms follow copies of the weights that are downloaded from mirrors rather than from Google.[1]
Section 2.2 grants the core permission: "You may use, reproduce, modify, Distribute, perform or display any of the Gemma Services only in accordance with the terms of this Agreement". The terms do not separate commercial from non-commercial use and contain no cap on users or revenue.[1] That distinguished them from Meta's Llama licenses, which require companies with more than 700 million monthly active users to obtain a separate license.[11] Section 3.1 also lets a redistributor add its own intellectual property statement and set "additional or different terms" for its modifications or for a Model Derivative as a whole, provided those terms do not conflict with the agreement.[1]
Model Derivatives
The definition of "Model Derivatives" has three parts: modifications to Gemma, works based on Gemma, and "any other machine learning model which is created by transfer of patterns of the weights, parameters, operations, or Output of Gemma, to that model in order to cause that model to perform similarly to Gemma". It names knowledge distillation through intermediate representations and training on synthetic data generated by Gemma as examples.[1] A model trained from scratch on Gemma-generated data, with the aim of matching Gemma's behavior, is therefore a derivative that carries the same terms, a point TechCrunch highlighted in its 2025 coverage.[11]
Redistribution obligations
The terms treat hosted access as distribution. "Distribution" covers any sharing of Gemma or a Model Derivative with a third party, "including by providing or making Gemma or its functionality available as a hosted service via API, web access, or any other electronic or remote means".[1] Anyone who reproduces or distributes Gemma or a Model Derivative must meet four conditions in Section 3.1:[1]
- Include the use restrictions of Section 3.2 "as an enforceable provision" in any agreement that governs use or distribution of the model, and tell downstream users that the model is subject to those restrictions.
- Give every third-party recipient a copy of the agreement.
- Make modified files carry prominent notices that they were modified.
- For distributions other than through a hosted service, include a "Notice" text file with the sentence "Gemma is provided under and subject to the Gemma Terms of Use found at" the terms' web address (ai.google.dev/gemma/terms in copies archived through at least early November 2026, gemma.ai/resources/terms in the November 2026 text).
Together with the definition of Model Derivatives, the first condition makes the use restrictions travel with every redistributed copy, fine-tuned variant and distilled model.[1] Licenses in the OpenRAIL family use a similar propagation mechanism for their behavioral restrictions.
Prohibited Use Policy
Section 3.2 bars use of the Gemma Services for the uses set out in the Prohibited Use Policy, which is "hereby incorporated by reference", and in violation of applicable laws and regulations.[1] The policy applies to "Gemma or Model Derivatives" and covers using them oneself or allowing others to use them.[2]
| Category | Examples given in the policy |
|---|---|
| Rights violations | Content that infringes, misappropriates or violates rights, including copyright, including the models' own outputs |
| Dangerous, illegal or malicious activities | Child sexual abuse material; instructions for illegal substances, goods or services; encouraging crimes; violent extremism; unlicensed practice of legal, medical, accounting or financial professions; spam, scams, phishing and malware; circumventing safety filters |
| Harm to individuals or groups (listed under dangerous, illegal or malicious activities) | Content promoting hatred, harassment, violence or self-harm; generating personally identifying information for distribution; tracking people without consent; unfair impacts on protected characteristics; inferring sensitive personal information without required consents |
| Misinformation and deception | Claiming generated content was made by a human in order to deceive; impersonation without disclosure; misleading claims of expertise in health, finance, government services or law; automated decisions affecting material rights in finance, legal, employment, healthcare, housing, insurance and social welfare; defamation |
| Sexually explicit content | Pornography and sexual gratification, such as sexual chatbots; content for scientific, educational, documentary or artistic purposes is excluded |
Source: Gemma Prohibited Use Policy, last modified February 21, 2024. The policy has four top-level items; the harm items are listed under the second.[2]
Section 3.2 bars unlawful use separately, and several policy items are written without reference to legality. The ban on "Making automated decisions in domains that affect material or individual rights or well-being" names finance, legal, employment, healthcare, housing, insurance and social welfare as examples, and the sexual content clause bars generating sexually explicit content in general, with the stated exceptions.[1][2]
Remote restriction, termination and other provisions
Section 3.2 ends with a reservation of power to Google: "To the maximum extent permitted by law, 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."[1] The agreement does not say how such a restriction would be carried out. Because "Gemma Services" is defined to include Model Derivatives and hosted services, the clause on its face also covers models and services that third parties build and run themselves.[1]
Section 4.5 lets Google terminate the agreement if a licensee is in breach. On termination the licensee "must delete and cease use and Distribution of all copies of Gemma and Model Derivatives" in its possession or control. The definitions, the eligibility clause, the output clause and Sections 4.2 to 4.9 survive termination.[1] Other provisions are conventional: the agreement grants no rights in Google's trademarks (4.2), disclaims warranties for the models and their outputs (4.3), limits Google's liability (4.4), and is governed by California law, with exclusive jurisdiction in the state and federal courts of Santa Clara County (4.6).[1]
Treatment of outputs
Section 3.3 says: "Google claims no rights in Outputs you generate using Gemma. You and your users are solely responsible for Outputs and their subsequent uses."[1] The definitions state that outputs are not Model Derivatives, so text or images generated with Gemma do not themselves carry the redistribution conditions. Two limits remain. Using those outputs to train a model "to perform similarly to Gemma" turns that model into a Model Derivative, and the Prohibited Use Policy still governs what content may be generated, its first item covering "the outputs or results generated by Gemma or Model Derivatives".[1][2]
Models covered
Before March 2025 the terms applied to whatever models Google identified on its Gemma website. Since then they have named the covered models in an appendix.[4][6] The November 2026 appendix lists the following:[1][6][7]
| Appendix entry | AI Wiki article | First listed |
|---|---|---|
| Gemma 1, Gemma 1.1 | Gemma | March 2025 appendix |
| Gemma 2 | Gemma 2 | March 2025 appendix |
| Gemma 2 JPN | Gemma 2 | March 2025 appendix |
| Gemma 3 | Gemma 3 | March 2025 appendix |
| Gemma 3n | Gemma 3n | By May 2025 |
| PaliGemma, PaliGemma 2 | PaliGemma | March 2025 appendix |
| ShieldGemma, ShieldGemma 2 | ShieldGemma | March 2025 appendix |
| CodeGemma, CodeGemma 1.1 | CodeGemma | March 2025 appendix |
| RecurrentGemma | RecurrentGemma | March 2025 appendix |
| Gemma Scope | Gemma Scope | March 2025 appendix |
| DataGemma RIG, DataGemma RAG | None | March 2025 appendix |
| Gemma-APS | None | March 2025 appendix |
| T5Gemma | None | By July 2025 |
| EmbeddingGemma | EmbeddingGemma | By September 2025 |
| VaultGemma | None | By September 2025 |
| FunctionGemma, T5Gemma 2 | None | By December 2025 |
| TranslateGemma | None | By January 2026 |
Gemma 4 is not in the appendix, and neither are later releases such as DiffusionGemma and EmbeddingGemma 2, whose weights are also under Apache 2.0. MedGemma and TxGemma are not listed either; they use the Health AI Developer Foundations terms.[1][16]
Health AI Developer Foundations terms
Health AI Developer Foundations (HAI-DEF) is a Google program that offers "free open-weight models and open-source companion tools" for healthcare applications, among them MedGemma, MedSigLIP, TxGemma and several earlier foundation models.[29] MedGemma is "built on Gemma 3", and TxGemma "is based on the Gemma 2 family".[30][31] These models are distributed under the Health AI Developer Foundations Terms of Use (last modified November 15, 2024) and a matching Prohibited Use Policy (last modified November 14, 2024).[14][15] The MedGemma model card states that "The use of MedGemma is governed by the Health AI Developer Foundations terms of use", and the Hugging Face repositories for MedGemma and TxGemma are tagged with that license or "other" rather than the Gemma license.[16]
The HAI-DEF terms follow the structure and much of the wording of the Gemma terms, including the Model Derivatives definition, the redistribution conditions, the remote restriction clause and a prohibited use policy whose list closely tracks Gemma's. They add provisions for medical use and shift more risk to the user:[14][15]
| Provision | Gemma Terms of Use | HAI-DEF Terms of Use |
|---|---|---|
| Covered models | Models listed in the appendix | Models "identified at developers.google.com/health-ai-developer-foundations" |
| Extra use restriction | None | No use "that could cause a Health Regulatory Authority to deem Google to be a 'manufacturer' of a medical device" |
| Extra redistribution duty | None | Seek Health Regulatory Authorization "when applicable" |
| Outputs | Google claims no rights | Google "will not claim ownership over any original Outputs" and may generate the same or similar outputs for others |
| Source code | Not addressed | Recipe, inference, training and utility code under Apache 2.0 |
| Medical disclaimer | None | Google is not "furnishing any medical advice or health care services" |
| Remote restriction | Yes | Yes, "without limiting any of Google's other rights" |
| Indemnification | None | The user must defend and indemnify Google against claims arising from a breach, injury caused by use of the models, gross negligence, or infringement by the user's modifications and Model Derivatives |
| Termination | On breach | On breach, or if Google reasonably determines that a regulator may deem it a "manufacturer" of a medical device because of the user's use |
Gated access
Google distributes models under the Gemma terms through gated repositories. On Hugging Face, the Gemma 3 repository tells visitors that "you have to accept the conditions to access its files and content" and that they are required to review and agree to Google's usage license; the repositories carry the Hub's "gemma" license tag.[12] Google's fine-tuning tutorials instruct users to accept the license with the "Agree and access repository" button and then authenticate with a Hugging Face token.[13] On Kaggle, Google's setup guide said that before first use "you must request access to the model through Kaggle", accepting "the Gemma use policy and license terms" through a consent form.[20] The MedGemma repositories are gated in the same way on the HAI-DEF terms.[16]
The agreement itself does not depend on the gate: it is accepted by use and defines Gemma "regardless of the source that you obtained it from".[1] Gemma 4 repositories are not gated and are tagged apache-2.0.[24]
Open source status
The Open Source Definition maintained by the Open Source Initiative requires that a license "must not discriminate against any person or group of persons" and "must not restrict anyone from making use of the program in a specific field of endeavor".[17] The OSI's Open Source AI Definition asks for the freedom to "Use the system for any purpose and without having to ask for permission", along with access to code and information about training data.[18] The Gemma terms forbid listed uses, require those prohibitions to be imposed on every downstream user, and let Google restrict use, so they do not fit either definition. Google has not presented them as an open source license: it called Gemma an "open model" in 2024, and when it announced Gemma 4 it described those models as "the first in the Gemmaverse to be released under the OSI-approved Apache 2.0 license".[10][21]
Outside reviewers reached the same conclusion. A June 2024 Nature report on a study by the Radboud University language scientist Mark Dingemanse and his colleague Andreas Liesenfeld, which assessed 40 large language models described as open, said that models including Llama and Google DeepMind's Gemma were "in fact, just 'open weight'". A Google spokesperson told Nature that the company chose to label Gemma as open rather than open source.[19] Florian Brand, an AI researcher at Trier University, told TechCrunch in 2025 that licenses such as Gemma's and Llama's "cannot reasonably be called 'open source.'"[11] Models released under such terms are commonly described as open-weight models.[19]
Criticism
Prominent criticism followed the release of Gemma 3 in March 2025. TechCrunch reported on March 14, 2025 that developers on X had complained that the license made commercial use of Gemma 3 "a risky proposition", and that smaller firms worried that Google could "pull the rug" on their businesses by asserting the more onerous clauses. The article called Gemma's license "generally less burdensome" than Llama's but singled out the remote restriction clause.[11]
The people TechCrunch quoted raised several distinct objections:[11]
- Nick Vidal, head of community at the Open Source Initiative, said that "the actual terms impose various legal and practical hurdles that deter businesses" from building so-called open models into their products.
- Florian Brand said that "Most companies have a set of approved licenses, such as Apache 2.0, so any custom license is a lot of trouble and money", and that the threat of enforcement deters adoption even though developers had not enforced their terms aggressively.
- Han-Chung Lee, director of machine learning at Moody's, said custom licenses such as Gemma's make the models "not usable" in many commercial scenarios.
- Eric Tramel, a staff applied scientist at Gretel, said that "Model-specific licenses make specific carve-outs for model derivatives and distillation, which causes concern about clawbacks", and that Gemma 3 seemed a solid release that "the market can't adopt" because of its license structure.
- Yacine Jernite, head of machine learning and society at Hugging Face, called on providers such as Google to move to open license frameworks and to "collaborate more directly" with users on broadly accepted terms.
The article did not include a response from Google.[11] When Gemma 4 moved to Apache 2.0, The Register summarized the old arrangement as a license that "had prohibited use of the models in certain scenarios and reserved the right to terminate a user's access", and wrote that enterprises could now deploy the models "without fear of Google pulling the rug out from under them".[23]
Gemma 4 and the move to Apache 2.0
Google announced Gemma 4 on April 2, 2026 under the Apache License 2.0; its release log dates the first Gemma 4 weights to March 31.[21][22][32] In the launch post on Google's blog, the section on licensing opens "You gave us feedback, and we listened." and says the license "provides a foundation for complete developer flexibility and digital sovereignty".[22] A companion post on the Google Open Source Blog said that "developers need autonomy, control, and clarity in licensing", and that applying standard Apache terms gave developers clarity "without the need to navigate prescriptive terms of service".[21] Neither post mentioned the Prohibited Use Policy or the remote restriction clause by name.[21][22]
Google did not relicense the earlier models. The Gemma Terms of Use were revised with an April 1, 2026 date to state that they cover the models in the appendix, with a link to the Gemma 4 license, and the terms archive page sends readers to the latest Gemma Terms of Use for Gemma models and Model Derivatives, and to the Apache license for Gemma 4.[1][3] As of November 2026, Hugging Face repositories for earlier releases such as Gemma 3 still carry the "gemma" license tag and the access gate, while Gemma 4 repositories are ungated and tagged apache-2.0.[12][24] A model fine-tuned from Gemma 3 therefore remains a Model Derivative under the Gemma terms even after Gemma 4's release.[1]
Third-party models under the Gemma terms
Because the obligations attach to Model Derivatives, models that other organizations build on Gemma weights carry the Gemma terms alongside whatever license their authors choose.
An example outside language modeling is Physical Intelligence's π0, a vision-language-action model for robot control. Its paper says the vision-language backbone weights "are initialized from PaliGemma".[26] The company's openpi repository, which publishes π0, π0-FAST and π0.5 checkpoints, is licensed under Apache 2.0 at the top level, and on September 30, 2025 a file named LICENSE_GEMMA.txt containing the February 21, 2024 text of the Gemma Terms of Use was added to it in a commit titled "Added the Gemma license".[25] The LeRobot ports of π0 and π0.5 on Hugging Face declare "license: gemma" and list google/paligemma-3b-pt-224 as their base model.[27]
Some derivatives inherit more than one license. The Gemma-2-Llama-Swallow models released by tokyotech-llm, built by continual pre-training of Gemma 2 models and instruction-tuned on synthetic data built for Japanese, list both the Gemma Terms of Use and the Llama 3.3 Community License in their model cards.[28]
Continue reading
- Apache License 2.0The Apache License, Version 2.0 is a permissive open source software license published by the Apache Software Foundation (ASF) in January 2004 and approved by the ASF board on January 21, 2004.Mentioned twice above
- OpenRAILOpenRAIL (Open Responsible AI Licenses, also written Open RAIL) is a family of licenses for machine learning models, source code, data and applications that gives royalty-free permission to use, modify and…Listed in See also
- Open Source AI DefinitionThe Open Source AI Definition (OSAID) is a document published by the Open Source Initiative (OSI) that states what an artificial intelligence system must make available, and under what termsListed in See also
- Open-Weight LLM License ComparisonAs of November 2026, the most widely used licenses that put no user, revenue or use conditions on commercial use of open-weight LLMs are two standard, OSI-approved ones: Apache-2.0 (Alibaba's Qwen3, most of…Listed in See also
- List of open-weight modelsThis is a list of open-weight models: artificial intelligence models whose trained parameters (weights) the developer has published for download, so that the model can be run, studied and in most cases…Listed in See also
- OpenwashingOpenwashing (also written open-washing or open washing) is the practice of presenting a product, organisation or technology as "open" while withholding the things that openness normally implies.Listed in See also
More to read
Topics:AI Policy & RegulationLarge Language ModelsOpen Source AI
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28Google. "Gemma Terms of Use". Last modified April 1, 2026. ai.google.dev/...terms (redirects to gemma.ai/...terms). Accessed November 17, 2026.
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Google. "Gemma Prohibited Use Policy". Last modified February 21, 2024. ai.google.dev/...prohibited_use_policy (redirects to gemma.ai/...prohibited-use). Accessed November 17, 2026.
- ^1 ^2 ^3Google. "Terms of Use archive". Gemma documentation. ai.google.dev/...terms-archive (redirects to gemma.ai/...terms-archive). Accessed November 17, 2026.
- ^1 ^2 ^3Google. "Gemma Terms of Use" (last modified February 21, 2024). ai.google.dev/...terms_02_21_24; as published on launch day, Internet Archive snapshot of February 22, 2024, web.archive.org/...terms
- ^Google. "Gemma Terms of Use" (last modified April 1, 2024). ai.google.dev/...terms_04_01_24
- ^1 ^2 ^3 ^4 ^5Google. "Gemma Terms of Use" (last modified March 24, 2025). Internet Archive snapshot of April 1, 2025. web.archive.org/...terms
- ^1 ^2 ^3Google. "Gemma Terms of Use", appendix as archived by the Internet Archive (first captures listing each later entry). web.archive.org/...terms ; web.archive.org/...terms ; web.archive.org/...terms ; web.archive.org/...terms ; web.archive.org/...terms ; web.archive.org/...terms
- ^1 ^2 ^3Google. "Gemma Terms of Use" (last modified April 1, 2026). Internet Archive snapshot of April 2, 2026. web.archive.org/...terms
- ^1 ^2 ^3Jeanine Banks and Tris Warkentin. "Gemma: Introducing new state-of-the-art open models". The Keyword (Google), February 21, 2024. blog.google/...gemma-open-models
- ^1 ^2 ^3Anne Bertucio and Helen King. "Building Open Models Responsibly in the Gemini Era". Google Open Source Blog, February 21, 2024. opensource.googleblog.com/...esponsibly-gemini-era
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Kyle Wiggers. "'Open' AI model licenses often carry concerning restrictions". TechCrunch, March 14, 2025. techcrunch.com/...en-carry-concerning-restrictions
- ^1 ^2Google. "google/gemma-3-27b-it" model repository (access conditions and license tag). Hugging Face. huggingface.co/...gemma-3-27b-it. Accessed November 17, 2026.
- ^Google. Full fine-tuning tutorial with Hugging Face Transformers. Gemma documentation. ai.google.dev/...huggingface_text_full_finetune (redirects to gemma.ai/...full-finetune). Accessed November 17, 2026.
- ^1 ^2 ^3Google. "Health AI Developer Foundations Terms of Use". Last modified November 15, 2024. developers.google.com/...terms
- ^1 ^2Google. "Health AI Developer Foundations Prohibited Use Policy". Last modified November 14, 2024. developers.google.com/...prohibited-use-policy
- ^1 ^2 ^3 ^4Google. "google/medgemma-1.5-4b-it" model card; "google/medgemma-4b-it" and "google/txgemma-9b-chat" repositories. Hugging Face. huggingface.co/...medgemma-1.5-4b-it. Accessed November 17, 2026.
- ^1 ^2Open Source Initiative. "The Open Source Definition". opensource.org/osd. Accessed November 17, 2026.
- ^Open Source Initiative. "The Open Source AI Definition 1.0". opensource.org/...open-source-ai-definition. Accessed November 17, 2026.
- ^1 ^2Elizabeth Gibney. "Not all 'open source' AI models are actually open: here's a ranking". Nature, June 19, 2024. nature.com/...d41586-024-02012-5
- ^Google. "Gemma setup" (last updated August 5, 2024). Internet Archive snapshot of December 21, 2025. web.archive.org/...setup
- ^1 ^2 ^3 ^4 ^5Nia Castelly, amanda casari and Olivier Lacombe. "Gemma 4: Expanding the Gemmaverse with Apache 2.0". Google Open Source Blog, April 2, 2026. opensource.googleblog.com/...averse-with-apache-20
- ^1 ^2 ^3Clement Farabet and Olivier Lacombe. "Gemma 4: Byte for byte, the most capable open models". The Keyword (Google), April 2, 2026. blog.google/...gemma-4
- ^Tobias Mann. "Google battles Chinese open-weights models with Gemma 4". The Register, April 2, 2026. theregister.com/...googles_gemma_4_open_weights
- ^1 ^2Google. "google/gemma-4-31B-it" model repository (license and access). Hugging Face. huggingface.co/...gemma-4-31B-it. Accessed November 17, 2026.
- ^Physical Intelligence. "openpi" repository: LICENSE (Apache 2.0) and LICENSE_GEMMA.txt, added in commit 2a34154, "Added the Gemma license", September 30, 2025. GitHub. github.com/...1547545855af53f395583a3a61d0166180a2
- ^Kevin Black et al. "π0: A Vision-Language-Action Flow Model for General Robot Control". arXiv:2410.24164, October 31, 2024. arxiv.org/...2410.24164
- ^LeRobot (Hugging Face). "lerobot/pi0_base" and "lerobot/pi05_base" model cards. Hugging Face. huggingface.co/...pi0_base ; huggingface.co/...pi05_base. Accessed November 17, 2026.
- ^tokyotech-llm. "Gemma-2-Llama-Swallow-9b-it-v0.1" model card. Hugging Face. huggingface.co/...Gemma-2-Llama-Swallow-9b-it-v0.1. Accessed November 17, 2026.
- ^Google. "Health AI Developer Foundations". Google for Developers. developers.google.com/...-ai-developer-foundations. Accessed November 17, 2026.
- ^Google. "MedGemma". Health AI Developer Foundations. developers.google.com/...medgemma. Accessed November 17, 2026.
- ^Google. "TxGemma model card". Health AI Developer Foundations. developers.google.com/...model-card. Accessed November 17, 2026.
- ^1 ^2Google. "Releases". Gemma documentation, gemma.ai. Accessed November 17, 2026. gemma.ai/...releases
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
1 revision · v1 · 4,387 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent verifier checked 104 claims against 41 sources and diffed about 100 archived copies of the terms; 1 material (HAI-DEF indemnity and termination) and 9 minor fixes applied.
Cite this page: AI Wiki. "Gemma Terms of Use." aiwiki.ai, updated 17 Nov 2026, fact-checked 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/gemma_terms_of_use