Open Weights

An AI system like a chat program consists at its core of a huge table of numbers. These numbers are produced during training, that is, while the system learns from millions of example texts. They are called weights because each number determines how strongly a signal is allowed to pass through the system’s computational pathway. Open weights means: the manufacturer makes exactly this file of numbers publicly available for download. Whoever has it can run the system on their own hardware, modify it, and redistribute it. The alternative are closed models, where the numbers reside only on the provider’s servers and one never gets to see them.

What the download shifts in terms of power

With a closed model, every request runs through the provider’s data centers. The provider can raise prices, disable features, or block certain topics. It also, in principle, sees what its users enter. For a hospital or a law firm this is a real problem, because they handle confidential data.

Open weights dissolve this dependency. A company can run the model in its own server room without a single byte leaving the premises. Researchers benefit as well: they can examine how the model works internally instead of merely observing its answers. This is what makes investigations into bias or security vulnerabilities possible in the first place.

The downside is discussed just as fiercely. A published model cannot be recalled. Safeguards meant to prevent unwanted answers can be removed again by anyone with a bit of computing power. Critics therefore warn of misuse for disinformation or fraudulent emails, while proponents consider the benefits for control and competition to be greater.

Why open does not mean the same as open source

The term is often confused with open source, meaning software whose blueprint is freely viewable and freely usable. Open weights are less than that. Usually only the result of training is published, not the path that led there. Which texts were used and with what settings the computations were run often remains secret. So one gets the finished cake, not the recipe.

Then there are the licenses, that is, the legal terms of use. Many providers do allow the download but forbid certain fields of application or require a contract once a certain number of users is exceeded. Genuine open-source licenses know no such restrictions. Experts therefore deliberately distinguish between open weights and open models in the full sense.

Technically, the download is only the beginning. The number files of large models are several dozen gigabytes in size and require powerful graphics cards. Smaller variants now run on a good laptop if the numbers are stored more coarsely, saving storage space. Fine-tuning is also widespread: an open model is further trained on one’s own data so that it masters specialized terminology or a particular tone.

Llama, Mistral, and the dispute over AI rules

The best-known examples come from Meta with the Llama series, from the French provider Mistral, and from Chinese companies such as Alibaba and DeepSeek. Such models are collected on platforms like Hugging Face, a kind of public warehouse for AI files. Hundreds of thousands of variants are now stored there, many of them fine-tuned by individuals.

In business news, open weights mainly come up as a strategic topic. Whoever gives away a strong model for free weakens competitors' pricing and turns their own format into the standard. At the same time, lawmakers are wrestling with the question of how strictly freely available models should be regulated. The EU’s AI Act provides for somewhat lighter obligations for open models in some respects, but explicitly does not exempt particularly powerful systems from them.

In everyday life one encounters the topic directly less often, but it lies behind many products. Some translation apps, coding assistants, and voice assistants in cars rely on open models because licensing costs are eliminated. The trend of running AI on one’s own device without an internet connection also only works with downloadable weights.