Jump to content

Open weights

From Wikipedia, the free encyclopedia

Open weights are the publicly released learned parameters of a trained artificial intelligence model, principally its weights and biases. In an artificial neural network, weights are numerical values that determine how strongly inputs contribute to the network's calculations. Publishing these values allows others to download and use the model, while permission to modify, fine-tune or redistribute it depends on its license. The term is commonly applied to large language models, diffusion models and other generative artificial intelligence.[1][2] Open weights is contrasted to open-source artificial intelligence, which considers the public release of not only parameters but the model's source code, training data, evaluation results, intermediate checkpoints and technical documentation.

Open weights AI is a major issue in geopolitics and artificial intelligence controversies, sometimes characterized as an AI arms race or AI Cold War between the US and China.[11] Broadly, models released by Chinese companies, such as DeepSeek, Alibaba Cloud, Moonshot AI and Z.ai, use an open weights framework, under more permissive software licenses like Apache or MIT. United States AI companies, including OpenAI, Anthropic, Google DeepMind, SpaceXAI, and Meta Superintelligence Labs, favor a proprietary software framework, especially for larger models.[12] These are in part national government policy decisions, because public access to AI technology has many consequences.[12] Some US politicians have called to restrict US public access to Chinese AI tools.[12]

Popular open-source artificial intelligence project categories include large language models (LLM), machine translation tools, and chatbots.[13] Debate over the benefits and risks of open-sourced AI involve a range of factors like security, privacy and technological advancement.[14][15][16][17] As of July 2026, the largest open weight frontier model is Kimi K3, developed by Moonshot AI, at 2.8 trillion parameters,[18] followed by Alibaba Cloud's Qwen3.8 at 2.4 trillion parameters.[19] Open weights releases outside China are led by US lab Thinking Machines Lab, Nvidia's Nemotron family, and French Mistral AI.[3][5][6]

Characteristics

[edit]

During training, a neural network adjusts its parameters to improve its output. Weights determine the contribution of inputs to weighted sums, while biases are additional learned values. A model described as having seven billion parameters contains roughly seven billion learned values.[20]

Model openness

[edit]

Releasing model weights may form part of a broader effort to make an artificial intelligence model more transparent and reproducible. The Model Openness Framework evaluates whether components such as the model's parameters, source code, training data, evaluation results, intermediate checkpoints and technical documentation are publicly available under an open license. Under the framework, releasing only the final weights represents a lower degree of openness than releasing the materials needed to study, modify and reproduce the model.[21]

Model cards

[edit]

A model card is a document accompanying a trained machine learning model that describes a model’s intended uses, limitations, training data and evaluation results. On Hugging Face, model cards are stored as Markdown README.md files with additional metadata in each model repository, helping users evaluate a model before downloading or deploying it.[22][23]

Comparison with open-source AI

[edit]

The Open Source Initiative distinguishes open weights from open-source artificial intelligence when the materials needed to study and modify the model's development process are not also available.[1]

Comparison of open weights and open-source artificial intelligence[1]
Component Open weights Open-source
Weights and biases Released Released
Training code Not shared Fully shared
Intermediate checkpoints Withheld Optional, but desirable
Training dataset Not shared or disclosed Released where legally permitted
Training-data composition Partially disclosed or undisclosed Fully disclosed

Benefits and concerns

[edit]

Open weights can support local deployment, customization, independent research and competition without requiring access through the original developer's API. However, widely distributed weights generally cannot be withdrawn and may be modified to remove safeguards. A 2024 report by the National Telecommunications and Information Administration examined these benefits and risks and recommended continued monitoring rather than immediate restrictions.[24][25]

Professors Peter Henderson and Mark Lemley argue that open-weight license terms are largely unenforceable because they presuppose intellectual property rights in ML model weights that do not exist. Whereas computer programs are typically copyrightable because they are "derived from a human-written copyrightable expression of an idea", model weights are generated automatically by a machine learning algorithm. The functional nature of ML model weights likely precludes copyright protection as well.[26] A whitepaper published by the Intellectual Property Owners Association in 2020 proposes a sui generis right in trained AI models.[27]

Open-weight large language models

[edit]

As of August 2026, the largest open weights models, with over a trillion parameters, are predominantly released by the AI industry in China, led by the larger Alibaba Cloud (Qwen3.8, 2.4T) and Deepseek (V4, 1.6T), and AI tigers including Moonshot AI (Kimi K3, 2.8T) and Z.ai (GLM-5.3, 753B). The largest models outside China are released by Thinking Machines Lab (Inkling, 975B), Nvidia (Nemotron 3 Ultra, 550B), and Mistral AI (Mistral Large 3, 675B).[3][5][6]

The following is a list of notable large language models and model families for which at least one version has been released with publicly accessible weights.[28][29]

See also

[edit]

References

[edit]
  1. 1 2 3 "Open Weights: not quite what you've been told". Open Source Initiative. Retrieved 28 July 2026.
  2. ↑ "Neural networks: Nodes and hidden layers". Google for Developers. Google. 3 December 2025. Retrieved 28 July 2026.
  3. 1 2 3 Piper, Kelsey (2024-02-02). "Should we make our most powerful AI models open source to all?". Vox. Retrieved 2025-12-16. Cite error: The named reference ":0" was defined multiple times with different content (see the help page).
  4. ↑ Cite error: The named reference s256 was invoked but never defined (see the help page).
  5. 1 2 3 "Trump, Xi and the AI race that could reshape global power". ABC News. 2026-05-11. Retrieved 2026-08-13. Cite error: The named reference ":1" was defined multiple times with different content (see the help page).
  6. 1 2 3 Chin, Josh; Huang, Raffaele (2025-11-11). "The AI Cold War That Will Redefine Everything". Wall Street Journal. ISSN 0099-9660. Retrieved 2026-08-13. Cite error: The named reference ":2" was defined multiple times with different content (see the help page).
  7. ↑ Wright, Webb (2026-07-01). "The AI Cold War Just Got Chillier". Gizmodo. Retrieved 2026-08-13.
  8. ↑ Guggenheim, Benjamin (2026-05-14). "AI & Tech Brief: The AI Cold War". The Washington Post. ISSN 0190-8286. Retrieved 2026-08-13.
  9. ↑ Thompson, Nicholas; Nicholas Thompson, Ian Bremmer (2018-10-23). "The AI Cold War That Threatens Us All". WIRED. Retrieved 2026-08-13.
  10. ↑ Kai-Fu Lee (September 25, 2018). AI Superpowers: China, Silicon Valley, and the New World Order. Boston, Mass: Houghton Mifflin. ISBN 9781328546395. OCLC 1035622189.
  11. ↑ [3][4][5][6][7][8][9][10]
  12. 1 2 3 McCabe, David; Isaac, Mike; Swanson, Ana; Conger, Kate (4 August 2026). "White House Whipsaws Silicon Valley (and Itself) Over A.I. Rules". The New York Times.
  13. ↑ Castelvecchi, Davide (29 June 2023). "Open-source AI chatbots are booming — what does this mean for researchers?". Nature. 618 (7967): 891–892. Bibcode:2023Natur.618..891C. doi:10.1038/d41586-023-01970-6. PMID 37340135.
  14. ↑ Sandbrink, Jonas (7 August 2023). "ChatGPT could make bioterrorism horrifyingly easy". Vox. Retrieved 14 November 2024.
  15. ↑ Toma, Augustin; Senkaiahliyan, Senthujan; Lawler, Patrick R.; Rubin, Barry; Wang, Bo (December 2023). "Generative AI could revolutionize health care — but not if control is ceded to big tech". Nature. 624 (7990): 36–38. Bibcode:2023Natur.624...36T. doi:10.1038/d41586-023-03803-y. PMID 38036861.
  16. ↑ Widder, David Gray; Whittaker, Meredith; West, Sarah Myers (November 2024). "Why 'open' AI systems are actually closed, and why this matters". Nature. 635 (8040): 827–833. Bibcode:2024Natur.635..827W. doi:10.1038/s41586-024-08141-1. ISSN 1476-4687. PMID 39604616.
  17. ↑ Davies, Pascale (20 February 2024). "What is open source AI and why is profit so important to the debate?". Euronews. Retrieved 28 November 2024.
  18. ↑ "China's Moonshot unveils world's largest open AI model, closing in on US rivals", Reuters, retrieved 2026-07-31
  19. ↑ "Alibaba unveils its largest AI model yet, DeepSeek's latest model is ultra-low cost", Reuters, retrieved 2026-08-03
  20. ↑ "Machine Learning Glossary: ML Fundamentals". Google for Developers. Google. Retrieved 28 July 2026.
  21. ↑ White, Matt; Haddad, Ibrahim; Osborne, Cailean; Liu, Xiao-Yang; Abdelmonsef, Ahmed; Varghese, Sachin Mathew; Le Hors, Arnaud (2024). "The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency, and Usability in Artificial Intelligence". arXiv:2403.13784 [cs.AI].
  22. ↑ "Model Cards". Hugging Face. Retrieved 28 July 2026.
  23. ↑ Mitchell, Margaret; Wu, Simone; Zaldivar, Andrew; Barnes, Parker; Vasserman, Lucy; Hutchinson, Ben; Spitzer, Elena; Raji, Inioluwa Deborah; Gebru, Timnit (2019). Model Cards for Model Reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency. Association for Computing Machinery. pp. 220–229. arXiv:1810.03993. doi:10.1145/3287560.3287596.
  24. ↑ "Dual-Use Foundation Models with Widely Available Model Weights Report". National Telecommunications and Information Administration. Retrieved 28 July 2026.
  25. ↑ "Open weights and American AI leadership". Microsoft. Retrieved August 6, 2026.
  26. ↑ Henderson, Peter; Lemley, Mark A. (June 2025). "The Mirage of Artificial Intelligence Terms of Use Restrictions". Indiana Law Journal. 100 (4): 1327–1387. arXiv:2412.07066.
  27. ↑ Tham, Yeen (2020-11-02). "Sui Generis Right for Trained AI Models" (PDF). Intellectual Property Owners Association. Retrieved 2026-08-08.
  28. ↑ Anel, Philippe (June 2026). "Open Weight Models". GitHub. Retrieved August 5, 2026.
  29. ↑ "AI Model Rankings". OpenRouter. OpenRouter, Inc. Retrieved September 7, 2026.
[edit]
Retrieved from "https://en.wikipedia.org/w/index.php?title=Open_weights&oldid=1374850294"