Open-weight model

An AI model learns from examples and stores what it has learned in huge tables of numbers. These numbers are called weights. They are the actual result of the learning process and can be many gigabytes in size. With an open-weight model, the maker makes these numbers publicly available for download. Whoever has the file can run the model on their own computer. The opposite are closed models: there, the numbers stay with the provider, and the model can only be reached through their servers.

Open weights are not the same as open source

The term is often confused with open source. With classic open-source software, the complete source code is openly available, and anyone may rebuild it. With an open-weight model, on the other hand, you only get the finished product. Which texts and images were used for training usually remains secret. The program that controlled the training is also rarely provided.

This is an important difference. Without the training data, no one from outside can check where certain abilities or errors come from. You can use the model, but not really trace its origins. Experts compare this to a cake: you get the finished cake, not the recipe. That is why this in-between form has earned its own name, open weights.

Then there are the licenses. Many well-known models may be used free of charge, but with conditions attached. Some forbid certain uses, others only allow use up to a certain size of one's own company. Freely downloadable, therefore, does not automatically mean freely usable.

What you can do with the downloaded weights

First, you need suitable hardware. Small models run on a good laptop, large ones need several specialized graphics cards. On top of that comes a program that loads the weights and carries out the computations. Such programs are usually free and set up within a few minutes.

The biggest advantage is customization. You can further train a model with your own examples so that it masters a particular subject area or a particular writing style better. This further training is called fine-tuning and is significantly cheaper than training from scratch. From a single open model, hundreds of variants often emerge, which other people then publish again.

A second advantage concerns data and control. If the model runs in-house, the inputs never leave the company network. For clinics, government agencies, or law firms, this is often decisive. In addition, the provider cannot shut down the model or secretly alter it. A file once downloaded stays exactly as it is.

Open models in the news and in products

Open weights regularly turn up in business news. Meta released its Llama models this way, the French company Mistral did something similar, and Chinese providers such as DeepSeek or Alibaba caused a stir with strong open models. Every such release tends to push down the prices that closed providers can charge.

Behind this also lies business strategy. Whoever gives away their model gains developers, attention, and influence over standards. Money is then made through cloud services, consulting, or additional products. At the same time, politics and research discuss the risks, because a public file cannot be recalled. Safeguards against misuse can be removed again by anyone with enough knowledge.

In everyday life, you often encounter such models without noticing. Translation features, chat assistants in apps, or search aids in company software are frequently based on a customized open model. Anyone who wants to experiment themselves can find thousands of models for download on platforms like Hugging Face. Getting a small language model running on your own computer today is a matter of an afternoon.