Open-Weight Models

An AI system like a chatbot consists at its core of a huge table of numbers. These numbers are called weights, and they are the result of a weeks-long learning process from text examples. In most well-known systems, these numbers remain a company secret: you can only use them via the provider’s servers. With open-weight models, it’s different. The manufacturer puts the number file online, usually several gigabytes in size, and anyone may download it. Anyone with a sufficiently powerful computer can then run the model completely offline.

Why open weights are changing the market

As long as the weights remain secret, every company and every government agency depends on the provider. It sets the price, the terms of use, and whether the model will even still exist next year. If it changes something, the customer has to adapt. Open weights break this dependency. Once a model has been downloaded, no one can shut it off anymore.

Then there’s data protection. Anyone running a model on their own machines doesn’t send a patient record or a draft contract to a foreign server. For hospitals, law firms, and government administrations, this is often the decisive point. Countries and universities also rely on open models for this reason.

Economically, open models push prices down. If a freely available system is nearly as good as an expensive subscription model, willingness to pay for the subscription declines. That’s exactly what has happened repeatedly since 2023. Analysts therefore watch very closely when a new open-weight model catches up to the commercial top tier.

What’s in the downloaded file – and what’s missing

The file contains the weights and a description of the architecture. This tells a program how to arrange and process the numbers. To run it, you need software like llama.cpp or Ollama, plus enough graphics memory. Small models run on a good laptop; large ones need special graphics cards costing tens of thousands of euros.

What’s almost always missing is the training data and the training code. You get the finished result, not the recipe. That’s why “open weight” is not the same as “open source.” With genuine open-source software, the complete blueprint is disclosed. With open weights, no one from outside can verify what texts the model was trained on.

A second pitfall is the license, i.e., the terms of use agreement. Some providers allow everything, others prohibit commercial use beyond a certain number of users or in certain application areas. “Freely downloadable” therefore doesn’t automatically mean “freely usable.” Checking the license file is a must before a company deploys such a model.

Well-known examples and typical uses

The best-known families are Llama from Meta, Mistral from France, Qwen from Alibaba, Gemma from Google, and the R1 series from DeepSeek. They are collected mainly on the Hugging Face platform, a kind of library for AI models. Hundreds of thousands of variants are hosted there, many of them adapted by individuals. This adaptation is called fine-tuning: you retrain a finished model with your own examples, for instance with legal texts.

In everyday life, you usually encounter open models invisibly. Many apps for translation, subtitles, or text correction use such a model in the background without mentioning it. On your own computer, you can install them with just a few clicks and query them without an internet connection.

In the news, they mainly appear in two contexts. First, in the race between China, the US, and Europe for technological independence. Second, in the debate about safety: a published model cannot be recalled, and built-in safeguards can be removed again during fine-tuning. Politics and research therefore argue over how open AI should be.