Open-weight model
In short
An open-weight model is an AI model whose trained weights are published for anyone to download and run on their own hardware. It is not always open source, because the licence, code or training data may still be restricted.
- Also called
- Open-weights model, Open model, Downloadable model, Model with widely available weights
- Topic
- The basics
- Level
- Beginner
- First described
- There is no single inventor of the idea. The US government set out the policy questions in detail in NTIA's July 2024 report on dual-use foundation models with widely available model weights.
A model's weights are the learned numbers that make it work. With an open-weight model, the developer publishes those numbers, so people can run the model offline, inspect it, or fine-tune it for their own task. Examples include Meta's Llama 3 family and OpenAI's gpt-oss models, released in August 202512.
Open-weight is not the same as open source. The Open Source Initiative's definition, published in October 2024, requires the freedoms to use, study, modify and share, plus enough information about the training data to rebuild a substantially equivalent system3. Many open-weight models do not meet that bar.
Licences vary a lot. OpenAI released gpt-oss under the permissive Apache 2.0 licence2. Meta's Llama 3.1 licence makes any company whose products had more than 700 million monthly active users when the model came out ask Meta for permission first4, and the Open Source Initiative says Llama's licence is not open source5.
Open weights bring trade-offs. A July 2024 US government report found that direct access to weights lets people strip built-in safety features more easily than with closed models, but concluded the evidence did not justify restricting such models at that time6.
The EU AI Act gives some relief to models released under a free and open-source licence with their weights published, by waiving certain documentation duties. The waiver does not apply to the most capable models judged to carry systemic risk7.
Example
A UK law firm that cannot send client documents to an outside AI company could download an open-weight model and run it on its own servers, so the documents never leave the building.
Why it matters
Open-weight models let businesses, researchers and hobbyists use capable AI without sending data to a provider or paying per message. They also make safeguards harder to enforce, which keeps them at the centre of AI policy debates.
Questions people ask
Is an open-weight model the same as open source?
Not necessarily. Open source AI, as defined by the Open Source Initiative, also requires free use for any purpose and information on training data3. Meta's Llama models publish weights, but the OSI says their licence is not open source5.
Are open-weight models free?
Usually free to download, but not always free of conditions. gpt-oss uses the permissive Apache 2.0 licence2, while Llama 3.1's licence requires very large companies to get Meta's permission4. Running a big model also needs costly hardware.
Can I run an open-weight model on my own computer?
Smaller ones, yes. OpenAI says gpt-oss-20b can run within 16GB of memory2. The largest open models need data centre chips.
Related terms
Tools that use it
- Meta Llama modelsMeta's family of downloadable large language models, from Llama 2 to the multimodal Llama 4
- Gemma 4Google's open-weight model family in four sizes for reasoning and agents on local hardware
- Olmo 3Ai2's fully open family of 7B and 32B language models, released with training data and code
- OllamaFree tool for running open AI models on a home computer, with optional paid cloud models
- GPT4AllFree desktop app from Nomic for running open-source AI models offline on a computer
Sources
- Grattafiori et al. (2024), The Llama 3 Herd of Models, arXiv:2407.21783
- OpenAI (2025), gpt-oss-120b and gpt-oss-20b Model Card, arXiv:2508.10925, with Hugging Face model page
- Open Source Initiative (28 October 2024), release of the Open Source AI Definition 1.0
- Meta, Llama 3.1 Community License Agreement, section 2 (Additional Commercial Terms)
- Open Source Initiative (February 2025), Meta's LLaMa license is still not Open Source
- NTIA, US Department of Commerce (July 2024), Dual-Use Foundation Models with Widely Available Model Weights
- Regulation (EU) 2024/1689 (AI Act), Article 53(2), via artificialintelligenceact.eu (EUR-Lex blocked automated access)