Download a model’s weights, run it on your own laptop, and it certainly feels open. Plenty of headlines call that “open-source AI.” The Open Source Initiative (OSI), the group that maintains the official definition of open source, would disagree in many cases. Open-weight AI models are now everywhere, but far fewer releases meet the formal bar for open source, and the difference changes what you are allowed to do with them.
What “Open Weights” Means
A trained model is mostly a very large file of numbers, called weights or parameters, that encode what it learned. An open-weight release lets anyone download those numbers and run the model without going through the company’s servers. That is useful: you can run it offline, keep your data on your own hardware, and fine-tune it for a narrow job.
What open weights do not guarantee is visibility into how the model was made. The training data, the code that processed it, and the recipe that produced the final numbers may all stay private. You get the finished product, not the kitchen.
What “Open Source AI” Requires
In October 2024 the OSI published version 1.0 of its Open Source AI Definition. It borrows the logic of traditional open-source software and asks that a system grant four freedoms: to use it for any purpose without asking permission, to study how it works, to modify it for any purpose, and to share it, with or without changes.
To make those freedoms real, the definition says the “preferred form” for making changes must be available. For a machine-learning system, that means three things:
-
Data information: a detailed description of the training data, including where it came from, how it was selected and filtered, and where the public portions can be obtained, even when the data itself cannot be shared.
-
Code: the complete source code used to train and run the system, covering data processing, training, validation, and inference.
-
Parameters: the weights themselves, under terms approved by the OSI.
Notice that weights are only one of the three requirements. A release that ships the weights alone, however generous, falls short of this definition.
Two Real Examples
Two well-known releases show how varied the “open” label can be. OpenAI published its gpt-oss models in August 2025 and, according to the company’s own announcement, released the weights under the Apache 2.0 license. The larger model is described as running on a single 80 GB GPU, and the smaller one on devices with 16 GB of memory. Apache 2.0 is a permissive license with few strings attached. Whether a given release also documents its training data and shares its training code is a separate question the announcement does not settle, so it is worth checking release by release.
Meta’s Llama family takes a different route. The Llama 3.1 community license allows broad use, but it adds conditions. Companies whose products had more than 700 million monthly active users when the model was released must request a separate license from Meta. Anyone distributing the model or a product built on it must display “Built with Llama,” and models trained using its outputs must carry a name beginning with “Llama.” Those are reasonable business terms, yet they limit use in ways that traditional open-source licenses do not.
Why the Label Matters to You
For most people, the practical differences come down to a few questions:
-
Can I run it myself? Open weights make local use possible. Our look at how on-device AI works explains why that can help with privacy and offline use.
-
Can I build a product on it? Check the license, not the headline. Some terms restrict large companies, require attribution, or limit how outputs may be used.
-
Can I verify it? Without training data information, outsiders cannot fully audit a model for bias, copyrighted material, or contamination from test questions.
-
Can I tweak it? Fine-tuning works with weights alone, which is why hobbyists and developers favor these releases. Tools like Ollama, covered in our open-source developer tools roundup, make downloading and running them simple.
It also helps to understand what the weights contain. Our explainer on how large language models work covers training and tokens, which makes it clearer why the data and code behind a model matter as much as the finished file.
A Quick Checklist Before You Trust the Word “Open”
When a new model is announced as open, read three things: the license text, the list of what was actually published (weights only, or data information and code too), and any usage restrictions. If a vendor says “open source,” compare its claim against the OSI definition above. If it says “open weights,” take that more modest phrase at face value, and judge the model on how well it performs for your own task.
FAQs
Is an open-weight model the same as open-source AI?
No. An open-weight model shares the trained parameters, while the OSI’s definition of open-source AI also asks for detailed data information and the code used to train and run the system, all under approved terms.
Can I use open-weight models commercially?
Often yes, but it depends on the license. Apache 2.0 is permissive, while custom licenses such as Meta’s Llama terms add conditions like attribution and a permission requirement for very large companies.
Do I need special hardware to run an open-weight model?
It depends on the size. OpenAI says its smaller gpt-oss model can run on devices with 16 GB of memory, while its larger one targets a single 80 GB GPU. Smaller community models can run on modest laptops.
Why does it matter whether the training data is disclosed?
Disclosure lets researchers check for bias, copyrighted material, and benchmark contamination. Without it, you are trusting the publisher’s description of what the model learned from.
Discussion about this post