Open Weights

An AI program like a chatbot consists at its core of a huge table of numbers. These numbers are called weights. They come into being during training, that is, in 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 run only 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 the model down. 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 premises. A clinic or a bank can run an open model on its own machines. Sensitive files then don’t go to some foreign corporation. That is often the real reason behind the decision, not the price.

Research benefits too. Scientists can take an open model apart and investigate why it gives certain answers. With closed systems, that’s 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 lies 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 helps little. What matters is the training data and the training procedure, and those usually remain secret.

That’s why open weights are more like a finished cake than a recipe. You get the result and may change it, decorate it, or cut it into pieces. How exactly it came about, you don’t learn. 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 allow free use only up to a certain size of one’s own company. Others prohibit certain use cases. So a model can be freely downloadable and still have strict rules. Anyone wondering what they’re allowed to do should read the license beforehand.

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. There, tens of thousands of variants are available that other developers have further refined.

Smaller open models now run on a good laptop or even on a phone. Some translation and dictation functions therefore work offline. Larger models, on the other hand, need expensive graphics cards and expertise to operate. The download is free, but ongoing operation is not.

In the news, open weights often come up in connection with politics. Governments are discussing whether freely available models are dangerous, because safety restrictions can be removed after the fact. At the same time, they are seen as a means against the market power of individual US corporations. Both arguments stand side by side within the same debate.