Open Weights
Open Weights means: A company releases the fully trained numerical values of its AI model for download. This lets anyone run and adapt the model themselves, but they don't have to learn what data it was trained on.
An AI program like a chatbot is, at its core, a huge table of numbers. These numbers are called weights. They emerge during training, that is, the learning phase in which the program evaluates millions of example texts. Everything the finished program can do is contained in these numbers. We speak of open weights when a company makes these numbers publicly available for download. Whoever has them and owns a sufficiently powerful computer can run the program without the company.
What the download shifts in terms of power
Most well-known AI systems only run on their providers' servers. You send a question over the internet and get an answer back. The provider can change prices at any time, tighten rules, or shut down the model. With open weights, this dependency disappears, because the downloaded file remains functional.
For companies, something else matters too: data doesn't have to leave the building. A clinic or a bank can run an open model on its own computers. Sensitive files then don't go to some outside corporation. That's often the real reason for the decision, not the price.
Research benefits as well. Scientists can take apart an open model and investigate why it gives certain answers. With closed systems, this is only possible from the outside, through trial and error.
Why open weights are not the same as open source
With classic free software, often called open source, the blueprint is laid open: the complete source code that programmers have written. You can read it and follow it line by line. With an AI model, the source code alone doesn't help much. What matters is the training data and the training process, and those usually remain secret.
That's why open weights are more like a finished cake than a recipe. You get the result and are allowed to change it, decorate it, or cut it into pieces. How exactly it came to be, you won't find out. Baking a model exactly the same way again would be impossible without the data.
On top of that come the licenses, that is, the legal terms of use. Many providers permit free use only up to a certain size of your own company. Others prohibit certain uses. So a model can be freely downloadable and still have strict rules. Anyone should read the license beforehand to know what's allowed.
Open models in products and headlines
The best-known examples are Meta's Llama models, the Mistral models from France, and several models from Chinese providers such as DeepSeek and Qwen. They are distributed via platforms like Hugging Face, a kind of collection point for AI models. Tens of thousands of variants are hosted there, further developed by other engineers.
Smaller open models now run on a good laptop or even on a phone. That's why some translation and dictation features work offline. Larger models, on the other hand, require expensive graphics cards and expertise to operate. The download is free, but running it is not.
In the news, open weights often come up in connection with politics. Governments discuss whether freely available models are dangerous because safety restrictions can be removed afterward. At the same time, they are seen as a means of countering the market power of individual US corporations. Both arguments exist side by side in the same debate.