open-weight model
An AI model whose learned parameters are published for anyone to download, run and modify, usually without the training data and code that produced them.
- Meta’s Llama 2, released under its own community license
- Mistral AI’s Mixtral 8x7B, under the Apache 2.0 license
- DeepSeek-R1, released under the MIT license in January 2025
- OpenAI’s gpt-oss-20b, which runs in 16 GB of memory
- Current meaning
- 2025
- The Open Source Initiative defines open weights and what they lack.
- Also called
- Open model
- The shorter form Mistral AI uses. NTIA says open foundation model.
- Opposite
- Closed model
- Weights kept by the maker; users send their data to its servers.
In plain words
In AI, an open-weight model is one whose trained weights anyone can download and run, while its training data and code usually stay private.
Most commercial chatbots run on closed models. The developer keeps the model on its own servers and sells access through an app or an API. The developer of an open-weight model publishes it instead. Its weights, the numbers a model learns in training, are posted as files that anyone can download. Examples include Meta’s Llama models, Mistral AI’s Mixtral, the reasoning model DeepSeek-R1 and OpenAI’s gpt-oss.
For a business, the appeal is control. The model can run on the company’s own hardware, so sensitive data does not go to an outside vendor. A startup is not tied to one provider.1 The company can also adapt the model to its own tasks through fine-tuning, which continues training on new examples.
Open weights are not the same as open source. The weights come with a license. Mixtral and gpt-oss use Apache 2.0, and DeepSeek-R1 uses MIT. Both are permissive licenses that allow commercial use. Meta called Llama 2 open source. Its license limits the largest companies and bans some uses, so the Open Source Initiative said it did not qualify.2 The initiative’s definition of open source AI also asks for the training code and detailed information about the training data.3 Most open-weight releases include neither.
Once weights are public, the developer cannot control every copy. In February 2023 Meta offered its first LLaMA model only to approved researchers. Within two weeks, a torrent of the model was posted on 4chan. In 2024 the US Commerce Department’s NTIA weighed wider participation in AI against misuse.4 It advised monitoring the risks rather than restricting such models.
How it works
A release usually includes the weights, code to run them and a license. Running the model, a step called inference, then happens on hardware the user chooses. Size sets the cost. OpenAI says its gpt-oss-120b runs on one graphics chip with 80 GB of memory, and gpt-oss-20b on devices with 16 GB.5
Holding the weights also means being able to change them. Fine-tuning a copy on new examples adapts it to a task, and the same step can strip out safety training. In 2023, researchers undid the safety training of Meta’s Llama 2-Chat models for less than $200 per model, using one graphics chip.6 Before releasing gpt-oss, OpenAI ran this test itself. It fine-tuned a version to answer biology and cybersecurity questions without refusing. OpenAI reported that the version did not reach the high capability level in its Preparedness Framework.5
What usually stays private is how the weights were made. The Open Source Initiative lists what open weights leave out: the training code, the training dataset and full details of the data. Without them, outsiders can run and adjust a model but cannot rebuild it or fully check what it learned from.
Origin and first known use
A compound of open, as in open source, and weights, the numbers a neural network learns in training. The term separates publishing the weights from publishing everything needed to rebuild a model. In 2023 Meta called Llama 2 open source, and the Open Source Initiative said its license did not qualify. That December, Mistral AI released Mixtral with open weights and called it an open-weight model. The initiative published its own definition of open weights in January 2025, and OpenAI used the term for gpt-oss that August.
- 2023 Meta releases LLaMA in four sizes under a noncommercial license. To prevent misuse, it grants access case by case to researchers in academia, government, civil society and industry.Meta AI, “Introducing LLaMA: A foundational, 65-billion-parameter large language model”, 24 February 2023 ai.meta.com →: source for 2023
- 2023 Meta releases Llama 2 under its own community license, which allows commercial use. Companies with over 700 million monthly active users must ask Meta for a license.Meta, “Llama 2 Community License Agreement”, version release date 18 July 2023 github.com →: source for 2023
- 2023 Mistral AI releases Mixtral 8x7B with open weights under the Apache 2.0 license. It says its mission is to deliver the best open models to developers.Mistral AI, “Mixtral of experts”, 11 December 2023 mistral.ai →: source for 2023
- 2024 NTIA’s report on open foundation models, those with widely available weights, finds too little evidence to decide on restrictions. It advises the US government to monitor the risks.National Telecommunications and Information Administration, “Dual-Use Foundation Models with Widely Available Model Weights Report”, 30 July 2024 ntia.gov →: source for 2024
- 2024 The Open Source Initiative publishes the Open Source AI Definition 1.0. It requires the parameters, the complete training and running code, and detailed information about the training data.Open Source Initiative, “The Open Source AI Definition – 1.0”, 28 October 2024 opensource.org →: source for 2024
- 2025 DeepSeek releases its R1 reasoning model and six smaller distilled models under the MIT license, inviting users to distill and commercialize them freely.DeepSeek, “DeepSeek-R1 Release”, 20 January 2025 api-docs.deepseek.com →: source for 2025
- 2025 The Open Source Initiative publishes a definition of open weights. It says weights released without training code and data fall short of open source AI.Open Source Initiative, “Open Weights: not quite what you’ve been told”, 29 January 2025 opensource.org →: source for 2025
- 2025 OpenAI releases gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license, its first language models with public weights since GPT-2.OpenAI, “Introducing gpt-oss”, 5 August 2025 openai.com →: source for 2025
In the wild
-
“Facebook’s Powerful Large Language Model Leaks Online”
-
“It is the strongest open-weight model with a permissive license and the best model overall regarding cost/performance trade-offs.”
Mistral AI, Mixtral announcement mistral.ai →: Mistral AI, Mixtral announcement, 11 Dec 2023Archived →: copy of Mistral AI, Mixtral announcement, 11 Dec 2023 -
“…we focus on open foundation models, defined here as those with broadly available model weights (e.g. Llama 2, Stable Diffusion XL).”
-
“…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…”
NTIA, US Department of Commerce, report ntia.gov →: NTIA, US Department of Commerce, report, 30 Jul 2024Archived →: copy of NTIA, US Department of Commerce, report, 30 Jul 2024 -
“Open-weight LLMs present unique challenges, as existing safeguards lack robustness to tampering attacks that modify model weights.”
Tamirisa et al., arXiv preprint arXiv 2408.00761 →: Tamirisa et al., arXiv preprint, 1 Aug 2024Archived →: copy of Tamirisa et al., arXiv preprint, 1 Aug 2024 -
“Open Weights refer to the final weights and biases of a trained neural network.”
Open Source Initiative, explainer opensource.org →: Open Source Initiative, explainer, 29 Jan 2025Archived →: copy of Open Source Initiative, explainer, 29 Jan 2025 -
“Open weights is the accepted term for when model weights of a language model are available on the internet for people to download.”
Nathan Lambert, on the Lex Fridman Podcast lexfridman.com →: Nathan Lambert, on the Lex Fridman Podcast, 3 Feb 2025Archived →: copy of Nathan Lambert, on the Lex Fridman Podcast, 3 Feb 2025 -
“Open-source and open-weight AI models are made freely available by developers for anyone in the world to download and modify.”
The White House, America’s AI Action Plan whitehouse.gov →: The White House, America’s AI Action Plan, 23 Jul 2025Archived →: copy of The White House, America’s AI Action Plan, 23 Jul 2025 -
“While we have made other models including Whisper and CLIP available openly, gpt-oss models are our first open-weight language models since GPT‑2.”
OpenAI, product announcement openai.com →: OpenAI, product announcement, 5 Aug 2025Archived →: copy of OpenAI, product announcement, 5 Aug 2025 -
“Open weights also give customers greater control.”
Not to be confused with
-
Open-source AI
Open-source AI, in the Open Source Initiative’s definition, also needs the complete training code and enough training-data detail to build an equivalent system. Most open-weight models publish only the weights and code to run them.
-
Free API access
A free chatbot or API sends each prompt to a model that stays on its maker’s servers, and the user never holds the weights. An open-weight model is a copy the user downloads and runs.
-
Weights
Weights are the numbers inside every trained model, open or closed. Open-weight describes a release decision: whether those numbers are published for others to download and run.
-
Foundation model
Foundation model describes how a model is built and reused: trained on broad data, then adapted to many tasks. Open-weight describes how it is released. A foundation model can be open-weight or closed.
Questions
What is the difference between open-weight and open-source AI?
The maker of an open-weight model publishes its trained weights, so anyone can download and run it. Open-source AI, as the Open Source Initiative defined it in October 2024, also requires the complete code used to train and run the system. It also requires enough information about the training data to build an equivalent system. Most open-weight releases leave out the training data and code, so they fall short of that definition.
Is Llama open source?
Meta called Llama 2 open source, but the Open Source Initiative says its license does not qualify. The Llama 2 license, dated 18 July 2023, requires companies with more than 700 million monthly active users to ask Meta for a license. It also restricts some uses through an acceptable use policy. Either way, its weights can be downloaded, so it is an open-weight model.
Can open-weight models be used commercially?
It depends on the license that comes with the weights. Mistral AI’s Mixtral 8x7B and OpenAI’s gpt-oss models use Apache 2.0, a permissive license that allows commercial use. DeepSeek released R1 under the MIT license and invited users to distill and commercialize it freely. Meta’s Llama 2 license allowed commercial use but required the largest companies to ask Meta first.
Why would a company choose an open-weight model over a closed one?
Mainly for control. An open-weight model can run on the company’s own hardware, so sensitive data does not have to go to an outside vendor. The White House’s 2025 AI Action Plan adds that startups can use such models without depending on a closed model provider. The company can also fine-tune the model for its own tasks. Smaller models need modest hardware: OpenAI’s gpt-oss-20b runs in 16 GB of memory.
Can safety training be removed from an open-weight model?
Yes. Whoever holds the weights can fine-tune them. In 2023, researchers undid the safety training of Meta’s Llama 2-Chat models for less than $200 per model, using one GPU. OpenAI tested a deliberately non-refusing version of gpt-oss before release, and researchers are building safeguards meant to resist such tampering. In 2024, NTIA advised the US government to monitor these risks rather than restrict such models.
Watch
Cite this entry
“Open-weight model.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Open-weight model. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Open-weight model.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/