Open model weights explained: what they are and why they matter
Model weights are the numerical values a language model learned during training, and they determine how it calculates an answer from an input. Open model weights means that the developer publishes these values as a file, so you can run the model on your own hardware, subject to the terms of its licence.
What model weights are technically
A language model consists of two parts: an architecture, meaning the calculation path, and the weights, meaning the very large number of parameters plugged into that path. Only both together make up the model. Whoever holds the weights can start the model, inspect it, adapt it further and move it to another location. Whoever only uses an interface sends requests to a model they never get to see.
Open model weights, open source and closed models compared
The three terms are often treated as the same, yet they mean different things.
| Feature | Open model weights | Open source AI model | Closed model via interface |
|---|---|---|---|
| Weights available | yes, as a file | yes | no |
| Training code and information on training data | not necessarily | usually yes | no |
| Usage rights | the developer's own licence | free licence in the open source sense | the provider's terms of use |
| Where it runs | your choice, including your own data centre | your choice | the provider's infrastructure |
| Changing the operator | possible, the weights move with you | possible | only by changing the model |
Open therefore does not automatically mean free. Many models with open weights are released under licences with their own rules, for example on commercial use, usage restrictions or the distribution of adapted versions. These terms differ from model to model and can change between versions. Check the model licence separately for each model and before productive use, ideally together with your legal department.
Why open model weights decide your exit options
Our industry report names portability of the model as one criterion of operational sovereignty, so that the dependency does not simply shift elsewhere. The background is a clear gap: 71 % of companies use US clouds, while only 8 % would prefer to (Bitkom Cloud Report 2026, published by Bitkom, the German digital industry association). If you use a closed model via an interface, you can only change provider by also changing the model. Instructions to the model, test routines and approvals then have to be validated again. With open weights the model stays the same and only the operator changes.
This applies especially to adapted models. When a model is retrained on your data, new weights are created. Clarify in the contract who owns them and in which format you receive them.
Practical example from our work
We prefer language models with open weights when we build AI in your own data centre. We justify the choice of model in writing during the initial assessment. If the collaboration ends, we hand over the model weights in the standard format for open models, together with data, configuration and documentation. The process is described on our page on switching AI provider. Where the weights are stored and who can access them is among the review criteria on our page on GDPR compliant AI.
