Open-Weight Model

At its core, an AI model consists of a huge table of numbers. These numbers are called weights, and they are the result of a weeks-long learning process based on a vast number of example texts. Whoever possesses these numbers possesses the finished model and can use it. With an open-weight model, the manufacturer makes exactly these numbers publicly available for download. You download a file of several gigabytes and let it run on your own computer. The counterpart is closed models: there, the numbers stay with the provider, and you may only ask questions through their website.

Why open weights are shifting the balance of power

Anyone using a closed model is entirely dependent on the provider. The provider can raise prices, disable features, or replace the model with a new version without warning. Companies that have built their product on top of it then face a problem. A downloaded model, by contrast, stays as it is. It still works even if the manufacturer loses interest or goes out of business.

The second major point is data protection. A hospital or a law firm may not simply send sensitive documents to a foreign server. An open-weight model can run in its own data center without a single byte leaving the premises. For many public authorities and companies in Europe, this is the decisive reason for even considering such models in the first place.

Finally, research benefits. Scientists can take an open model apart and investigate why it gives certain answers. With a closed model, outsiders are left guessing. Critics point out, however, that open weights can also be misused, since built-in safety restrictions can subsequently be removed again.

What exactly is disclosed – and what isn’t

The name is more precise than it sounds. Only the weights – that is, the learned numbers – are disclosed, plus usually a description of the model’s architecture. As a rule, the training data and the code used for training are not disclosed. So you get the finished cake, not the recipe and not the list of ingredients.

This is precisely why experts strictly distinguish between “open weight” and “open source.” Genuine open-source software discloses everything, so that you could reproduce the result yourself. With open-weight models, this is impossible, because nobody knows which texts went into it. Many providers nevertheless call their models “open source” – this is marketing and is regularly criticized.

The license also matters. Some models may be used commercially without restriction, others only up to a certain number of users or not for commercial purposes at all. Open weights are also practically useful for so-called fine-tuning: you train the finished model a bit further using your own data, thereby specializing it, for example for legal texts.

Llama, Mistral, and the race with China

The best-known examples come from Meta, the Facebook group, whose Llama models have been downloadable since 2023. Mistral comes from France, while models such as DeepSeek and Qwen come from China. When DeepSeek released a very strong open model in early 2025, stock markets reacted noticeably – the assumption that good AI is only possible with billion-dollar budgets and behind closed doors was shaken.

In everyday life, you usually encounter this indirectly. Translation features, voice assistants, or search helpers in apps often run on adapted open models, without it being labeled anywhere. Smaller variants now even fit on a laptop or a smartphone.

Anyone wanting to try it out themselves can find the files on platforms like Hugging Face, a kind of collection point for AI models. Programs such as Ollama or LM Studio launch a model on your own computer with just a few clicks. The answers are then usually somewhat weaker than with the large commercial services – but in return, everything runs offline and costs nothing except electricity.