Model Weights Release

At its core, an AI model consists of a huge table of numbers. These numbers are called weights, and they are the result of training, i.e. the model’s weeks-long learning phase. They determine which answer the program gives to an input. Normally, companies keep these numbers to themselves and only let users talk to the model over the internet. With a model weights release, however, the maker provides the numbers as a file for download. Anyone with suitable hardware can then run the model themselves.

What open weights change about the distribution of power

As long as the weights are secret, there is exactly one way to reach the model: through the provider’s servers. The provider sets the price, sees every request, and can shut down or change the model at any time. Anyone who builds their business on it is dependent on a single contractual partner. A release of the weights removes this dependency. The file then sits on one’s own hard drive and cannot be recalled.

This also matters for data protection. Hospitals, government agencies, and banks are often not allowed to send sensitive data to external data centers. With downloaded weights, everything stays in-house. Researchers benefit as well: they can look inside the model and investigate why it makes certain mistakes. With a closed model, all that’s left to them is guessing from the outside.

But there is a downside. Whoever has the weights can remove built-in safeguards again. A model that is supposed to refuse to give instructions for weapons or malware can be retrained with manageable effort. This is precisely the concern providers cite when they keep their strongest models closed. That’s why every major release is a political issue in the industry.

What’s really in the file

What gets released is usually a package of two things. First, the weights themselves, often ranging from several gigabytes to many hundreds of gigabytes in size. Second, a description of the architecture, so that a program can correctly assign the numbers. On top of that comes a license, i.e. a usage agreement. Some licenses permit everything, others prohibit commercial use or certain fields of application.

What’s important is what’s typically not included: the training data and the training code. So you get the finished cake, not the recipe. That’s why experts speak of “open weights” rather than “open source.” Genuine open-source software releases every component, so that one could rebuild the result oneself. For most released AI models, this is impossible.

With the weights, you can do more than just ask questions. They can be further trained, for example on the technical jargon of a professional field. They can be shrunk by storing the numbers more coarsely, thereby reducing memory requirements. A model that originally required several specialized graphics cards sometimes then runs on a good gaming PC. No one who only has access via the provider’s servers is allowed to make these modifications.

Releases in headlines and products

In the news, weights releases regularly come up when Meta, Mistral, Alibaba, or Chinese labs publish new models. Such announcements even move stock prices. That’s because a freely downloadable model that is almost as good as an expensive subscription model drives down prices across the entire market. Observers call this the commoditization of AI: what everyone gets for free is hard to sell.

In practice, one encounters released weights on platforms like Hugging Face, a kind of download portal for AI models. Tens of thousands of variants are hosted there, many of them modified by volunteers. Programs that let you run a language model offline on your own laptop also rely on such files. Within companies, they’re often invisibly embedded in products, for example in a customer service chatbot.

A common misconception is that open weights are automatically weaker than closed ones. The gap fluctuates and has often been small recently. A second misconception is that they’re always free to operate. The download is free, but the computing power is not. Anyone who runs a large model themselves on an ongoing basis pays for electricity and hardware and needs staff for it.