Open-weights model

An AI system like a chatbot consists at its core of a huge table of numbers. These numbers arise during learning from example texts and determine how the system responds. They are called the model’s weights. With an open-weights model, the maker publishes exactly these numbers for free download. Anyone with the right hardware can then run the model themselves, without asking the maker. The counterpart is closed models: there, the numbers stay in the company’s data center, and you can only reach them via their website or interface.

Why this matters

Open weights shift the balance of power in the industry. Companies and government agencies can run a model on their own servers. Sensitive data then never leaves the building, which is crucial for patient records or contracts. It also avoids creating a dependency on a single provider. A downloaded model still works even if the maker raises prices or shuts down the service.

For research, open weights are almost even more valuable. Only those allowed to look inside the model can investigate why it makes certain mistakes. Critics, on the other hand, point to risks: a published model cannot be recalled. Safety guardrails meant to prevent unwanted responses can later be removed again.

How it works

The maker uploads the weights as files to a platform, usually several gigabytes in size. Along with this comes a license, i.e. a usage agreement. It governs what is allowed, for example whether the model may be used commercially. Many licenses permit almost everything, while some exclude major competitors or certain applications.

It’s important to note the difference from open source in regular software. There, the readable program code is open, and every step can be traced. With open weights, you only get the result of the learning process. Which texts were used in training and which settings were applied usually remains secret. An analogy: you get the finished cake along with permission to resell it, but not the recipe. That’s why experts deliberately speak of open weights rather than open source.

Because the weights are modifiable, adapted versions quickly emerge from them. Users retrain the model with their own data so that it masters specialized language or a particular tone. Others shrink it so it can run on an ordinary laptop.

Where you encounter the term

The term comes up in the news when major providers release new models. Meta's Llama series, Mistral from France, and the Chinese models from DeepSeek or Alibaba are well-known examples. Such announcements even move stock prices, since a strong free model puts pressure on the prices of paid services.

The central hub for such files is the platform Hugging Face. Hundreds of thousands of models are available there for download. With programs like Ollama or LM Studio, a smaller model can be started on one’s own computer, with no internet connection required at all. The term also comes up in politics: in EU rules on artificial intelligence, there is debate over what obligations should apply to freely distributed models.