Maple

Open-weight policy

Who can run which AI model

A dated record of the laws, rules, licenses and official findings that decide how AI model weights are released, with one primary source per entry.

An open-weight model is one whose trained parameters, the weights, are published for anyone to download. A license then sets what you may do with them. Some documents call this “open source”; the Open Source Initiative’s definition asks for more than weights. This record also lists closed releases, license limits and findings that argue against wide release.

48 entriesLatest Checked

2026

8 entries

U.S. Department of Homeland Security Notice: DHS publishes an order excluding DeepSeek products from civilian procurement and requiring removal of DeepSeek products and third-party products developed using DeepSeek implementations.

Federal Register, 91 FR 66346 (Docket DHS-2026-0628) (primary source, opens in a new tab)

Details
On open weights Silent

The notice publishes an order by the Secretary of Homeland Security under the Federal Acquisition Supply Chain Security Act of 2018, 41 U.S.C. 1321-1328. It covers Hangzhou DeepSeek Artificial Intelligence Co., Ltd and affiliated entities as defined in 48 CFR 2.101. The signed order is reproduced as images in the notice’s appendix. The order in the appendix carries an original release date of July 24, 2026; the Federal Register notice publishing it is dated October 20, 2026.

Source type
Federal Register
Checked
Entry
#2026-10-20-dhs-deepseek-exclusion-order

Bureau of Industry and Security Notice: The Unified Agenda says a planned interim final rule will formally rescind the January 2025 AI diffusion rule, which BIS has stopped enforcing.

Federal Register, Introduction to the Unified Agenda, 2026-16603 (primary source, opens in a new tab)

Details
On open weights Silent

In the Bureau of Industry and Security regulatory-plan section, RIN 0694-AJ90, 'Implementation of the AI Action Plan Through Export Controls,' is described as an interim final rule that will rescind the January 2025 Framework for Artificial Intelligence Diffusion and establish a replacement framework. The notice does not publish the rescission itself.

Source type
Federal Register
Checked
Entry
#2026-08-14-bis-diffusion-rescission-agenda

European Parliament and Council Law: Regulation (EU) 2026/1744 sets high-risk AI Act dates of December 2, 2027 for Annex III and August 2, 2028 for Annex I.

Official Journal, Regulation (EU) 2026/1744 (primary source, opens in a new tab)

Details
On open weights In cases of observed or reported circumvention of the preventive measures or other safeguards, adequate corrective measures should be taken provided that such measures are reasonable, taking into account the specific AI system, including its release and distribution strategy (such as open-source releases).

The act is dated July 8, 2026 and enters into force on the third day after Official Journal publication. It replaces Article 113, third paragraph, points (a) and (c), and adds point (d). It does not replace point (b), the August 2, 2025 date for Chapter V. The Commission news page states that the omnibus enters into force on July 27, 2026.

Source type
Official Journal
Checked
Entry
#2026-07-27-ai-act-digital-omnibus

Bureau of Industry and Security Rule: BIS states it is enforcing the worldwide advanced-computing license requirement only for specified country groups, citing its May 13, 2025 guidance.

Federal Register, 2026-14132 (primary source, opens in a new tab)

Details
On open weights Silent

The final rule, published July 14, 2026, and effective July 10, 2026, removes the United Arab Emirates from Country Groups D:3 and D:4 and adds it to Country Group A:5. It says BIS is only enforcing the section 742.6(a)(6)(iii)(A) license requirement for destinations in Country Groups D:1, D:4, or D:5, with stated exclusions, and for entities headquartered in or with an ultimate parent headquartered in Country Group D:5 or Macau.

Source type
Federal Register
Checked
Entry
#2026-07-14-bis-advanced-computing-enforcement

Bureau of Industry and Security Guidance: BIS says its non-enforcement of the AI Diffusion Rule's new requirements does not cover exports for D:5 or Macau-headquartered entities.

BIS guidance, May 31, 2026 (primary source, opens in a new tab)

Details
On open weights Silent

The guidance says that in May 2025 BIS announced it would not enforce the AI Diffusion Rule's new compliance requirements. It says a license is still required under section 742.6(a)(6)(iii)(A) for specified advanced computing items when the recipient is headquartered in, or has an ultimate parent headquartered in, Country Group D:5 or Macau.

Source type
Government publication
Checked
Entry
#2026-05-31-bis-diffusion-enforcement-guidance

Center for AI Standards and Innovation Evaluation: CAISI evaluates the open-weight model DeepSeek V4 Pro and states its capabilities lag the frontier by about 8 months.

NIST, CAISI Evaluation of DeepSeek V4 Pro (primary source, opens in a new tab)

Details
On open weights In April 2026, the Center for AI Standards and Innovation (CAISI) evaluated the open-weight AI model DeepSeek V4 Pro ("DeepSeek V4").

The page says CAISI's evaluations, which include non-public benchmarks, indicate that DeepSeek V4 performs similarly to GPT-5, and that DeepSeek's self-reported results placed it nearer to Opus 4.6 and GPT-5.4. It says CAISI served the model from cloud-based H200 and B200 GPUs.

Source type
Government publication
Checked
Entry
#2026-05-01-caisi-deepseek-v4-pro

Google Release: Gemma 4 model weights under the Apache 2.0 license, in E2B, E4B, 26B MoE, and 31B sizes.

Google blog, Gemma 4 (primary source, opens in a new tab)

Details
On open weights That’s why Gemma 4 is released under a commercially permissive Apache 2.0 license.

The post says “We are releasing the Gemma 4 model weights” in E2B, E4B, 26B MoE, and 31B sizes. It also says “This open-source license provides a foundation” under the heading “An open-source license,” and: “They complement our Gemini models, giving developers the industry's most powerful combination of both open and proprietary tools.” The terms page, last modified April 1, 2026, still points models other than Gemma 4 to the Gemma Terms of Use.

License
Apache License, Version 2.0 (opens in a new tab)
Commercial use
each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
Weights
Downloadable
Previous license
Gemma Terms of Use (opens in a new tab)
Source type
Company page
Checked
Entry
#2026-04-02-gemma-4-apache

New York Law signed: Senate Bill S8828, chapter 96, repeals and replaces the RAISE Act text and defines frontier model and model weight.

New York Senate Bill S8828 (primary source, opens in a new tab)

Details
On open weights CYBERSECURITY PRACTICES TO SECURE UNRELEASED MODEL WEIGHTS FROM UNAUTHORIZED MODIFICATION OR TRANSFER BY INTERNAL OR EXTERNAL PARTIES;

The Senate actions list 'Mar 27, 2026' as 'signed chap.96.' The bill text defines a frontier model as a foundation model trained with more than 10^26 operations and a large frontier developer as one with annual gross revenue above $500 million. The sponsor memo calls it a chapter amendment to Chapter 699 of the Laws of 2025.

Source type
Legislature
Checked
Entry
#2026-03-27-new-york-raise-chapter-amendment

2025

20 entries

Governor Kathy Hochul Law signed: Governor Hochul signed the RAISE Act, S6953B/A6453B, with agreed chapter amendments requiring large frontier developers to publish safety protocols.

Governor Hochul press release, December 19, 2025 (primary source, opens in a new tab)

Details
On open weights Silent

The release says the agreed-upon chapter amendments require large AI developers to publish information about their safety protocols and to report incidents within 72 hours, and create an oversight office in the Department of Financial Services. It states penalties of up to $1 million for a first violation and up to $3 million for subsequent violations.

Source type
Government publication
Checked
Entry
#2025-12-19-new-york-raise-act-signed

Center for AI Standards and Innovation Evaluation: CAISI evaluates Kimi K2 Thinking, which it calls an open-weight model released November 6, 2025, and compares it with U.S. models.

NIST, CAISI Evaluation of Kimi K2 Thinking (primary source, opens in a new tab)

Details
On open weights Kimi K2 Thinking, an open-weight AI model released on November 6, 2025 by Moonshot AI.

The page says that at release Kimi K2 Thinking was the most capable model from a PRC-based developer in this evaluation and still lagged leading U.S. models, while its capabilities improved on the previous open-weight model frontier. It also says downloads a month after release were about 10 percent of DeepSeek R1's and under 5 percent of gpt-oss's.

Source type
Government publication
Checked
Entry
#2025-12-12-caisi-kimi-k2-thinking

European Commission Communication: The Commission’s December 5, 2025 communication sets out an explanatory notice and template for public summaries of general-purpose AI model training content.

European Commission, training-summary template, C(2025) 8311 (primary source, opens in a new tab)

Details
On open weights Providers of all general-purpose AI models placed on the Union market must fulfil the above obligation, including providers of general-purpose AI models released under free and open-source licenses, in so far as the models fall within the scope of the AI Act.

A footnote on the same page says the Article 53(2) exception for a free and open-source general-purpose AI model does not apply to the obligation to make the summary public. Paragraph (1) says the Chapter V rules apply as of August 2, 2025.

Source type
Government publication
Checked
Entry
#2025-12-05-gpai-training-summary-template

European Commission Report: The Commission’s guidelines for general-purpose AI model providers, C(2025) 7719 final, dated November 19, 2025, include an exception for models released under a free and open-source license.

European Commission, GPAI guidelines, C(2025) 7719 (primary source, opens in a new tab)

Details
On open weights These transparency-related obligations do not apply to providers of general-purpose AI models that are released under a free and open-source licence, under certain conditions, unless the model is a general-purpose AI model with systemic risk.

The guidelines say Article 53(2) and Article 54(6) except some duties for models released under a free and open-source license whose parameters, including the weights, are public, unless the model has systemic risk. Copyright policy and the training-content summary stay in place. The text also restates the August 2, 2025, August 2, 2026, and August 2, 2027 dates. The Commission’s library page lists the guidelines with “Publication 18 July 2025” and a Communication approving “the content of the draft Guidelines.”

Source type
Government publication
Checked
Entry
#2025-11-19-gpai-provider-guidelines

OpenAI Release: gpt-oss-safeguard-120b and gpt-oss-safeguard-20b weights under the same Apache 2.0 license as gpt-oss.

OpenAI, Introducing gpt-oss-safeguard (primary source, opens in a new tab)

Details
On open weights Today, we’re releasing a research preview of gpt-oss-safeguard, our open-weight reasoning models for safety classification tasks, available in two sizes: gpt-oss-safeguard-120b and gpt-oss-safeguard-20b. These models are fine-tuned versions of our gpt-oss open models and available under the same permissive Apache 2.0 license, allowing anyone to use, modify, and deploy them freely. Both models can be downloaded today from Hugging Face.

The post describes the two models as fine-tuned versions of the gpt-oss models and says they are under the same Apache 2.0 license. The license file quoted here is the LICENSE in openai/gpt-oss, opened with the August 5, 2025 release.

License
Apache License, Version 2.0 (opens in a new tab)
Commercial use
each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
Weights
Downloadable
Previous license
None recorded
Source type
Company page
Checked
Entry
#2025-10-29-gpt-oss-safeguard

Center for AI Standards and Innovation Evaluation: CAISI reports on DeepSeek models it describes as open-weight, including jailbreaking results and a comparison with U.S. models.

CAISI, Evaluation of DeepSeek AI Models (primary source, opens in a new tab)

Details
On open weights To evaluate DeepSeek models, which are available as open-weight models, CAISI downloaded their model weights from the model sharing platform Hugging Face and deployed the models on CAISI's own cloud-based servers.

The report evaluates DeepSeek R1, R1-0528, and V3.1 against GPT-5, GPT-5-mini, gpt-oss, and Opus 4. It says DeepSeek models were much more likely to follow malicious hijacking instructions than the evaluated U.S. frontier models, and that the evaluated U.S. open-weight model matched or exceeded DeepSeek on that measure.

Source type
Government publication
Checked
Entry
#2025-09-30-caisi-deepseek-evaluation

California Law signed: Senate Bill 53 was approved and chaptered as Chapter 138, the Transparency in Frontier Artificial Intelligence Act.

California Business and Professions Code, Chapter 25.1 (primary source, opens in a new tab)

Details
On open weights Cybersecurity practices to secure unreleased model weights from unauthorized modification or transfer by internal or external parties.

The code says Chapter 25.1 was added by Stats. 2025, Ch. 138, Sec. 2 (SB 53), effective January 1, 2026. It defines a frontier model as a foundation model trained using more than 10^26 operations and defines 'model weight' as a numerical parameter adjusted through training.

Source type
Legislature
Checked
Entry
#2025-09-29-california-sb-53-chaptered

State Council Plan: The State Council issues an opinion that supports AI open-source communities and opening models, tools, and datasets.

State Council, Opinions on the AI Plus action, Guofa [2025] No. 11 (primary source, opens in a new tab)

Details
On open weights 支持人工智能开源社区建设，促进模型、工具、数据集等汇聚开放，培育优质开源项目。

Guofa [2025] No. 11 is dated August 21, 2025 and was posted on August 26, 2025. A separate paragraph says China will promote making AI technology open-source and accessible in international cooperation. The opinion also says safety assessment and filing rules for AI should be improved.

Source type
Government publication
Checked
Entry
#2025-08-26-china-ai-plus-open-source

OpenAI Release: gpt-oss-120b and gpt-oss-20b weights under the Apache 2.0 license, described as open-weight language models.

OpenAI, Introducing gpt-oss (primary source, opens in a new tab)

Details
On open weights We’re releasing gpt-oss-120b and gpt-oss-20b—two state-of-the-art open-weight language models that deliver strong real-world performance at low cost. Available under the flexible Apache 2.0 license

The post says “gpt-oss models are our first open-weight language models since GPT-2” and that the weights are “freely available for download on Hugging Face.” On the same page: “Once an open-weight model is released, adversaries may be able to fine-tune the model for malicious purposes.” It says that even with robust fine-tuning, maliciously fine-tuned models “were unable to reach high capability levels” under the Preparedness Framework, and that “These findings informed our decision to release the gpt-oss models.”

License
Apache License, Version 2.0 (opens in a new tab)
Commercial use
each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
Weights
Downloadable
Previous license
None recorded
Source type
Company page
Checked
Entry
#2025-08-05-gpt-oss

The White House Plan: The AI Action Plan says open-source and open-weight models can be downloaded and modified, and that a release decision is up to the developer.

Winning the AI Race: America's AI Action Plan (primary source, opens in a new tab)

Details
On open weights Open-source and open-weight AI models are made freely available by developers for anyone in the world to download and modify.

The section 'Encourage Open-Source and Open-Weight AI' also says the decision of whether and how to release an open or closed model is fundamentally up to the developer. A recommended action assigns the Commerce Department, through NTIA, to convene stakeholders on adoption by small and medium-sized businesses.

Source type
Government publication
Checked
Entry
#2025-07-23-ai-action-plan-open-models

European Commission Report: The Commission publishes a voluntary code of practice for providers of general-purpose AI models under the AI Act.

European Commission, The General-Purpose AI Code of Practice (primary source, opens in a new tab)

Details
On open weights Silent

The page says the code was published on July 10, 2025, and that the Commission and the AI Board have confirmed it is an adequate voluntary tool. The Transparency and Copyright chapters address Article 53. The Safety and Security chapter addresses Article 55 models with systemic risk.

Source type
Government publication
Checked
Entry
#2025-07-10-gpai-code-of-practice

Department of Commerce Announcement: Commerce instructs staff not to enforce the AI Diffusion Rule and says BIS plans a regulation formalizing rescission.

Department of Commerce press release, May 13, 2025 (primary source, opens in a new tab)

Details
On open weights Silent

The release, headed for immediate release on May 13, 2025, says the AI Diffusion Rule was issued on January 15, 2025, with compliance requirements set for May 15, 2025. It says Under Secretary Jeffery Kessler instructed enforcement officials not to enforce the rule, and that BIS plans to publish a regulation formalizing the rescission and a replacement rule.

Source type
Government publication
Checked
Entry
#2025-05-13-bis-diffusion-rule-nonenforcement

Qwen Team Release: Qwen3 MoE and dense model weights under the Apache 2.0 license, including Qwen3-235B-A22B.

Qwen blog, Qwen3 (primary source, opens in a new tab)

Details
On open weights We are open-weighting two MoE models: Qwen3-235B-A22B, a large model with 235 billion total parameters and 22 billion activated parameters, and Qwen3-30B-A3B, a smaller MoE model with 30 billion total parameters and 3 billion activated parameters. Additionally, six dense models are also open-weighted, including Qwen3-32B, Qwen3-14B, Qwen3-8B, Qwen3-4B, Qwen3-1.7B, and Qwen3-0.6B, under Apache 2.0 license.

The post lists Qwen3-235B-A22B among the models it says are open-weighted under the Apache 2.0 license. On June 7, 2024, Qwen2-72B had remained on the Qianwen License. The LICENSE file opened for this quote is on Qwen/Qwen3-8B. Its appendix reads “Copyright 2024 Alibaba Cloud” and “Licensed under the Apache License, Version 2.0.”

License
Apache License, Version 2.0 (opens in a new tab)
Commercial use
each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
Weights
Downloadable
Previous license
Tongyi Qianwen LICENSE AGREEMENT (opens in a new tab)
Source type
Company page
Checked
Entry
#2025-04-29-qwen3

Meta Release: Llama 4 Scout and Maverick weights for download; the same post says it is not yet releasing Llama 4 Behemoth.

Meta AI blog, Llama 4 (primary source, opens in a new tab)

Details
On open weights We’re introducing Llama 4 Scout and Llama 4 Maverick, the first open-weight natively multimodal models with unprecedented context length support and our first built using a mixture-of-experts (MoE) architecture.

The post says Scout and Maverick are “available for download today” on llama.com and Hugging Face, and: “While we’re not yet releasing Llama 4 Behemoth as it is still training, we’re excited to share more technical details about our approach.” It describes Behemoth as 288B active parameters, 16 experts, and nearly two trillion total parameters. The license states “Llama 4 Version Effective Date: April 5, 2025,” and the 700 million clause is keyed to “the Llama 4 version release date.” The acceptable-use policy is https://www.llama.com/llama4/use-policy.

License
LLAMA 4 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Commercial use
If, on the Llama 4 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
Weights
Downloadable
Previous license
LLAMA 3.3 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Source type
Company page
Checked
Entry
#2025-04-05-llama-4

Cyberspace Administration of China Rule: Four Chinese agencies issue measures on labels for AI-generated synthetic content, effective September 1, 2025.

CAC, Measures for Labeling AI-Generated Content (primary source, opens in a new tab)

Details
On open weights Silent

The notice is dated March 7, 2025 and posted March 14, 2025. The measures apply to network information service providers in the situations covered by the algorithm-recommendation rules, the deep-synthesis rules, and the generative AI service measures. They require explicit and implicit labels on generated or synthetic text, images, audio, video, and virtual scenes.

Source type
Government publication
Checked
Entry
#2025-03-14-china-ai-content-labels

France and India Statement: France and India say they will pursue research on large language models that are open and freely reusable.

Élysée, Franco-Indian declaration on artificial intelligence (primary source, opens in a new tab)

Details
On open weights pour approfondir les partenariats de recherche pour développer des modèles de langage larges, ouverts et librement réutilisables afin de soutenir et promouvoir la diversité linguistique

The same declaration says the two countries will support shared resources including open-source tools, and that a framework should include free and open resources for countries, innovators, researchers, and citizens. The page is the French text of the joint declaration.

Source type
Government publication
Checked
Entry
#2025-02-12-france-india-open-models

Networking and Information Technology Research and Development (NITRD) National Coordination Office (NCO), National Science Foundation Request for information: On behalf of OSTP, NITRD asks for AI Action Plan comments and lists open source development among the topics.

Federal Register, 90 FR 9088 (primary source, opens in a new tab)

Details
On open weights Responses can address any relevant AI policy topic, including but not limited to: hardware and chips, data centers, energy consumption and efficiency, model development, open source development, application and use

The notice was published February 6, 2025, by the NITRD National Coordination Office at the National Science Foundation, on behalf of OSTP. Comments were due March 15, 2025. It cites Executive Order 14179 of January 23, 2025, and says Executive Order 14148 of January 20, 2025, revoked Executive Order 14110.

Source type
Federal Register
Checked
Entry
#2025-02-06-ostp-ai-action-plan-rfi

DeepSeek License change: DeepSeek-R1 code and models are released under the MIT License.

DeepSeek API docs news, 250120 (primary source, opens in a new tab)

Details
On open weights Code and models are released under the MIT License: Distill & commercialize freely!

The same news page says “Fully open-source model & technical report,” “DeepSeek-R1 is now MIT licensed for clear open access,” “Open for the community to leverage model weights & outputs,” and “6 small models fully open-sourced.” The MIT file on the R1 repository begins “Copyright (c) 2023 DeepSeek.” The DeepSeek-V3 LICENSE-MODEL file remained the October 23, 2023 DeepSeek License Agreement when fetched on October 16, 2026.

License
MIT License (opens in a new tab)
Commercial use
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software
Weights
Downloadable
Previous license
DEEPSEEK LICENSE AGREEMENT (opens in a new tab)
Source type
Company page
Checked
Entry
#2025-01-20-deepseek-r1-mit

Bureau of Industry and Security Rule: BIS adds ECCN 4E091 for certain closed-weight model weights and states it is not imposing controls on open-weight models.

Federal Register, Framework for Artificial Intelligence Diffusion, 90 FR 4544 (primary source, opens in a new tab)

Details
On open weights Accordingly, BIS and its interagency partners are not today imposing controls on open-weight models.

The interim final rule, 90 FR 4544, is effective January 13, 2025, with compliance for most changes on May 15, 2025. Its text says ECCN 4E091 does not control the parameters of a model that has been published as defined in section 734.7(a) of the EAR, and that no open-weight model was known to have been trained on more than 10^26 operations.

Source type
Federal Register
Checked
Entry
#2025-01-15-bis-eccn-4e091-diffusion-rule

2024

14 entries

DeepSeek Release: DeepSeek-V3 weights in the Hugging Face repository under the DeepSeek License Agreement, Version 1.0, dated October 23, 2023.

DeepSeek API docs, Introducing DeepSeek-V3 (primary source, opens in a new tab)

Details
On open weights Fully open-source models & papers

The file on the DeepSeek-V3 repository, fetched October 16, 2026, is still the DeepSeek License Agreement, Version 1.0, October 23, 2023, the same text as on DeepSeek-LLM. It is not the MIT License. The repository lists 163 safetensors shards. Use-based restrictions in Attachment A must be passed downstream, and section 7 reserves a right to restrict usage that violates the license.

License
DEEPSEEK LICENSE AGREEMENT (opens in a new tab)
Commercial use
a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model.
Weights
Downloadable
Previous license
None recorded
Source type
Company page
Checked
Entry
#2024-12-25-deepseek-v3

LF AI & Data Definition: Model Openness Framework 1.0 defines Class I, Class II, and Class III, and defines an open license.

LF AI & Data, Model Openness Framework Specification (primary source, opens in a new tab)

Details
On open weights An open license is a license allowing unrestricted usage, study, modification, and redistribution for any purpose.

The specification is titled as Version 1.0, dated Dec 17, 2024, from LF AI & Data Generative AI Commons, and it says “This is a stable document.” It says: “The MOF defines three classes (Class III: Open Model, Class II: Open Tooling Model, and Class I: Open Science Model).” Class III includes model architecture and final model parameters under type-appropriate open licenses. For model parameters it lists preferred CDLA-Permissive-2.0 and acceptable OSI-approved or open data licenses.

Source type
Standards body
Checked
Entry
#2024-12-17-model-openness-framework

Meta Release: Llama 3.3 weights under the Llama 3.3 Community License Agreement, which includes a 700 million monthly active user clause.

llama-models LICENSE, Llama 3.3 (primary source, opens in a new tab)

Details
On open weights including machine-learning model code, trained model weights, inference-enabling code, training-enabling code, fine-tuning enabling code

The license states “Llama 3.3 Version Release Date: December 6, 2024.” The 700 million monthly active user clause remains. The acceptable-use policy linked from the license is https://www.llama.com/llama3_3/use-policy. The llama-models README table lists Llama 3.3 as 12/04/2024.

License
LLAMA 3.3 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Commercial use
If, on the Llama 3.3 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
Weights
Downloadable
Previous license
LLAMA 3.2 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Source type
License file
Checked
Entry
#2024-12-06-llama-3-3

Open Source Initiative Definition: The Open Source AI Definition 1.0 requires parameters under OSI-approved terms, plus data information and code.

Open Source Initiative, Open Source AI Definition (primary source, opens in a new tab)

Details
On open weights An Open Source AI is an AI system made available under terms and in a way that grant the freedoms to: Use the system for any purpose and without having to ask for permission. Study how the system works and inspect its components. Modify the system for any purpose, including to change its output. Share the system for others to use with or without modifications, for any purpose.

The definition page says the preferred form of making modifications includes data information, code, and parameters, and that “Parameters shall be made available under OSI-approved terms.” It also says: “Open Source models and Open Source weights must include the data information and code used to derive those parameters.”

Source type
Standards body
Checked
Entry
#2024-10-28-osi-open-source-ai-definition

Governor Gavin Newsom Veto: Governor Newsom returned Senate Bill 1047 unsigned, a bill on safeguards for developers of large artificial intelligence models and a Board of Frontier Models.

Governor Newsom, SB 1047 veto message (primary source, opens in a new tab)

Details
On open weights Silent

The letter is dated September 29, 2024, and begins 'I am returning Senate Bill 1047 without my signature.' It says the bill would require developers of large artificial intelligence models, and those providing computing power to train them, to put safeguards in place, and would establish a Board of Frontier Models.

Source type
Government publication
Checked
Entry
#2024-09-29-california-sb-1047-veto

Meta Release: Llama 3.2 weights under the Llama 3.2 Community License Agreement, which includes a 700 million monthly active user clause.

llama-models LICENSE, Llama 3.2 (primary source, opens in a new tab)

Details
On open weights including machine-learning model code, trained model weights, inference-enabling code, training-enabling code, fine-tuning enabling code

The license states “Llama 3.2 Version Release Date: September 25, 2024.” No clause banning use of outputs to improve another large language model was found in this file. The 700 million monthly active user clause remains. The acceptable-use policy linked from the license is https://www.llama.com/llama3_2/use-policy. The llama-models README table lists Llama 3.2 and Llama 3.2-Vision on the same date.

License
LLAMA 3.2 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Commercial use
If, on the Llama 3.2 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
Weights
Downloadable
Previous license
LLAMA 3.1 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Source type
License file
Checked
Entry
#2024-09-25-llama-3-2

National Telecommunications and Information Administration Report: NTIA states that current evidence is not sufficient to determine that restrictions on open weight models are warranted, or that they never will be.

NTIA, Dual-Use Foundation Models with Widely Available Model Weights Report (primary source, opens in a new tab)

Details
On open weights current evidence is not sufficient to definitively determine either that restrictions on such open weight models are warranted, or that restrictions will never be appropriate in the future.

The report page is dated July 30, 2024, and says the report responds to section 4.6 of the October 2023 executive order after a February 2024 request for comment that drew 332 comments. It recommends monitoring and keeping open the possibility of later limits on model-weight availability.

Source type
Government publication
Checked
Entry
#2024-07-30-ntia-open-model-weights-report

Meta License change: Llama 3.1 weights for download; the license allows Llama outputs, including the 405B, to be used to improve other models.

Meta AI blog, Introducing Llama 3.1 (primary source, opens in a new tab)

Details
On open weights We’ve also made changes to our license, allowing developers to use the outputs from Llama models—including the 405B—to improve other models.

The license states “Llama 3.1 Version Release Date: July 23, 2024.” The ban on using outputs to improve another large language model is not in this file. Section 1.b.i allows use of the materials or outputs “to create, train, fine tune, or otherwise improve an AI model” when the name starts with “Llama” and “Built with Llama” is displayed. The 700 million monthly active user clause remains. The post says the models are available for download on llama.meta.com and Hugging Face.

License
LLAMA 3.1 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Commercial use
If, on the Llama 3.1 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
Weights
Downloadable
Previous license
META LLAMA 3 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Source type
Company page
Checked
Entry
#2024-07-23-llama-3-1-license-change

European Parliament and Council Law: Regulation (EU) 2024/1689 is published in the Official Journal, including rules on AI systems released under free and open-source licenses.

Official Journal, Regulation (EU) 2024/1689 (primary source, opens in a new tab)

Details
On open weights This Regulation does not apply to AI systems released under free and open-source licences, unless they are placed on the market or put into service as high-risk AI systems or as an AI system that falls under Article 5 or 50.

The act is dated June 13, 2024. The journal issue is dated July 12, 2024. Article 113 says it enters into force on the twentieth day after publication and applies Chapter V, on general-purpose AI models, from August 2, 2025, except Article 101. Article 111(3) gives providers of general-purpose AI models placed on the market before August 2, 2025 until August 2, 2027 to comply.

Source type
Official Journal
Checked
Entry
#2024-07-12-eu-ai-act-published

Qwen Team License change: Qwen2-72B remains under the Qianwen License, and Qwen2-0.5B, 1.5B, 7B, and 57B-A14B move to Apache 2.0.

Qwen blog, Qwen2 (primary source, opens in a new tab)

Details
On open weights This time, we change the licenses of our models to different ones. While Qwen2-72B as well as its instruction-tuned models still uses the original Qianwen License, all other models, including Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, and Qwen2-57B-A14B, turn to adopt Apache 2.0!

The license object above is the Apache 2.0 file on Qwen2-7B. It does not cover Qwen2-72B. The post says Qwen2-72B and its instruction-tuned models still use the original Qianwen License. That file’s 100 million monthly active user clause is the one quoted on the August 3, 2023 Tongyi Qianwen entry. The Qwen2-7B LICENSE appendix reads “Copyright 2024 Alibaba Cloud.”

License
Apache License, Version 2.0 (opens in a new tab)
Commercial use
each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
Weights
Downloadable
Previous license
Tongyi Qianwen LICENSE AGREEMENT (opens in a new tab)
Source type
Company page
Checked
Entry
#2024-06-07-qwen2-license-split

Mistral AI Release: Codestral under the Mistral AI Non-Production License for non-production use, with Apache 2.0 releases continuing.

Mistral AI news, MNPL (primary source, opens in a new tab)

Details
On open weights Codestral is released today under this new license

The May 29, 2024 post says: “This license allows developers to use our technology for non-commercial purposes and to support research work.” It also says: “Note that we will keep releasing models and code under Apache 2.0 as we progressively consolidate two families of products released under Apache 2.0 and the MNPL.” The license limits use to “testing, research, Personal, or evaluation purposes in Non-Production Environments,” and its definition of the model includes “instructed checkpoints, parameters.”

License
Mistral AI Non-Production License (opens in a new tab)
Commercial use
You shall not supply the Mistral Models or Derivatives in the course of a commercial activity, whether in return for payment or free of charge, in any medium or form, including but not limited to through a hosted or managed service (e.g. SaaS, cloud instances, etc.), or behind a software layer.
Weights
Downloadable
Previous license
Apache License, Version 2.0 (opens in a new tab)
Source type
Company page
Checked
Entry
#2024-05-29-mistral-mnpl

Meta Release: Llama 3 weights under the Meta Llama 3 Community License Agreement, including a 700 million monthly active user clause.

Meta AI blog, Meta Llama 3 (primary source, opens in a new tab)

Details
On open weights Today, we’re introducing Meta Llama 3, the next generation of our state-of-the-art open source large language model.

The license states “Meta Llama 3 Version Release Date: April 18, 2024.” It keeps the clause barring use of the materials or their output “to improve any other large language model (excluding Meta Llama 3 or derivative works thereof).” The acceptable-use policy linked from the license is https://llama.meta.com/llama3/use-policy. Attribution in the license uses “Built with Meta Llama 3.”

License
META LLAMA 3 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Commercial use
If, on the Meta Llama 3 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
Weights
Downloadable
Previous license
LLAMA 2 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Source type
Company page
Checked
Entry
#2024-04-18-llama-3

National Telecommunications and Information Administration Request for comment: NTIA asks how to define widely available model weights, which the notice calls open foundation models, and what rules might apply.

Federal Register, 89 FR 14059 (primary source, opens in a new tab)

Details
On open weights Dual use foundation models with widely available weights (referred to here as open foundation models) could play a key role in fostering growth among less resourced actors, helping to widely share access to AI's benefits.

The notice, Docket No. 240216-0052, RIN 0660-XC060, implements section 4.6 of the October 30, 2023 executive order. Comments were due March 27, 2024. The notice also says widely available weights may pose risks, including affirmative misuse and lack of monitoring.

Source type
Federal Register
Checked
Entry
#2024-02-26-ntia-widely-available-weights-rfc

Google Release: Gemma 2B and Gemma 7B model weights, with terms the post says permit commercial usage and distribution for all organizations.

Google blog, Gemma open models (primary source, opens in a new tab)

Details
On open weights Terms of use permit responsible commercial usage and distribution for all organizations, regardless of size.

The post says: “We’re releasing model weights in two sizes: Gemma 2B and Gemma 7B.” It calls them “open models.” The terms page incorporates the Gemma Prohibited Use Policy at ai.google.dev/gemma/prohibited_use_policy. Distribution is to carry the notice “Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.”

License
Gemma Terms of Use (opens in a new tab)
Commercial use
You may use, reproduce, modify, Distribute, perform or display any of the Gemma Services only in accordance with the terms of this Agreement
Weights
Downloadable
Previous license
None recorded
Source type
Company page
Checked
Entry
#2024-02-21-gemma

2023

6 entries

The President Order: Section 4.6 directs Commerce, through NTIA, to seek comment and report on dual-use foundation models whose model weights are widely available.

Federal Register, Executive Order 14110, 88 FR 75191 (primary source, opens in a new tab)

Details
On open weights When the weights for a dual-use foundation model are widely available--such as when they are publicly posted on the internet--there can be substantial benefits to innovation, but also substantial security risks, such as the removal of safeguards within the model.

Executive Order 14110 is dated October 30, 2023, and was published in the Federal Register on November 1, 2023. Section 4.6 gives the Secretary 270 days to solicit public input and submit a report to the President.

Source type
Federal Register
Checked
Entry
#2023-10-30-eo-14110-widely-available-weights

Department for Science, Innovation and Technology Report: A DSIT discussion paper for the 2023 AI Safety Summit describes open-source generative AI and access to model development.

GOV.UK, generative AI risks to 2025, Annex B (primary source, opens in a new tab)

Details
On open weights Open-source generative AI is facilitating rapid proliferation and increasing democratisation of generative AI by reducing the barriers to entry for developing models.

Paragraph 4 says open-source models had mostly lagged frontier models and were highly unlikely to be more capable than leading commercial frontier models by 2025, and that proliferation brings safety and security implications where safeguards are not effective. The parent page says the paper does not represent government policy.

Source type
Government publication
Checked
Entry
#2023-10-25-uk-open-source-generative-ai

Mistral AI Release: Mistral 7B under the Apache 2.0 license, which the post says can be used without restrictions.

Mistral AI news, Announcing Mistral 7B (primary source, opens in a new tab)

Details
On open weights We're releasing Mistral 7B under the Apache 2.0 license, it can be used without restrictions.

The September 27, 2023 post is the release notice and includes a download link. The Hugging Face repo mistralai/Mistral-7B-v0.1 lists license apache-2.0 and weight files, and it has no LICENSE file among its siblings.

License
Apache License, Version 2.0 (opens in a new tab)
Commercial use
each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
Weights
Downloadable
Previous license
None recorded
Source type
Company page
Checked
Entry
#2023-09-27-mistral-7b

Alibaba Cloud Release: Tongyi Qianwen materials, including trained model weights, under a license with a 100 million monthly active user clause.

Qwen-7B LICENSE, Tongyi Qianwen LICENSE AGREEMENT (primary source, opens in a new tab)

Details
On open weights consisting of trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code

The file states “Tongyi Qianwen Release Date: August 3, 2023” and names Alibaba Cloud. It also says: “You can not use the Materials or any output therefrom to improve any other large language model (excluding Tongyi Qianwen or derivative works thereof).” Materials include “trained model weights.”

License
Tongyi Qianwen LICENSE AGREEMENT (opens in a new tab)
Commercial use
If you are commercially using the Materials, and your product or service has more than 100 million monthly active users, You shall request a license from Us. You cannot exercise your rights under this Agreement without our express authorization.
Weights
Downloadable
Previous license
None recorded
Source type
License file
Checked
Entry
#2023-08-03-qwen-tongyi-license

Meta Release: Llama 2 weights for research and commercial use under the Llama 2 Community License Agreement, which includes a 700 million monthly active user clause.

Meta AI blog, Llama 2 (primary source, opens in a new tab)

Details
On open weights We’re now ready to open source the next version of Llama 2 and are making it available free of charge for research and commercial use. We’re including model weights and starting code for the pretrained model and conversational fine-tuned versions too.

The license is headed “Llama 2 Version Release Date: July 18, 2023.” It also says: “You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Llama 2 or derivative works thereof).” An Acceptable Use Policy is at https://ai.meta.com/llama/use-policy.

License
LLAMA 2 COMMUNITY LICENSE AGREEMENT (opens in a new tab)
Commercial use
If, on the Llama 2 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee's affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
Weights
Downloadable
Previous license
None recorded
Source type
Company page
Checked
Entry
#2023-07-18-llama-2

Cyberspace Administration of China Rule: Seven Chinese agencies publish interim measures for generative AI services offered to the public in China, effective August 15, 2023.

CAC, Interim Measures for Generative AI Services (primary source, opens in a new tab)

Details
On open weights Silent

The order, signed July 10, 2023 and posted July 13, 2023, says the measures apply to services that use generative AI to provide text, images, audio, or video to the public in China. Article 17 requires a security assessment and algorithm filing for services with public-opinion attributes or social-mobilization capacity. Article 22 defines a provider to include offering the service through a programmable interface.

Source type
Government publication
Checked
Entry
#2023-07-13-china-generative-ai-measures

All entries: 48 shown

Free to use

Cite this data

Free to quote, chart and republish under CC BY 4.0. All we ask is a credit with a link back. Sources keep their own terms. CC BY 4.0

Source: Maple, Open-weight AI policy (maple.ai/open-weights), data through 2026-10-20.

Latest entry · Checked

Download the data: data.json · data.csv · How entries are chosen

Maple runs open-weight models inside a hardware-isolated enclave. Maple cannot see your chats. How it works

Every entry cites one primary source, quotes the document’s own words about open weights, and is checked against that source before it publishes. Corrections keep the old wording visible.
Methodology · Corrections · Suggest an entry or a correction

RSS · JSON · CSV
Also kept by Maple: the privacy timeline