What are those "weights"?
During training, an AI model learns billions of numbers, which are the weights. Together they form the model's "brain": everything it learned is captured in those numbers. Whoever has the weights has the working model and can run it.
Open or closed?
- Open-weights (e.g. Llama, Mistral, Gemma): you can download the weights and host the model yourself, wherever you want.
- Closed models (e.g. GPT by OpenAI, Claude by Anthropic): you never get the weights. You use the model only through their API, which means your data goes to their servers.
Note: open-weights is not the same as fully "open source". With open-weights you get the working weights, but not necessarily the training data or all the code used to build the model. To use it, that is more than enough.
Why does this matter?
Because you hold the weights, you can put an open-weights model on your own or European servers and process your data without it ever going to an external provider. With a closed API model, you have to send your data to the vendor by definition. For data that must stay within the EU, that is the difference.
The trade-off
The largest closed models are often still more capable and easier to use today. Open-weights models, on the other hand, give you full control and privacy, but you handle the hosting and the compute yourself. The right choice depends on your sector, your data sensitivity and your budget.
Related terms
- EU hosting & zero retention: this is exactly where a self-hostable model makes the difference for your data.
- RAG: lets a model, open or closed, answer based on your own sources.
- Vector store: where your own data lives so the model can consult it.
