The landscape of language models is currently shifting in several places at once. Models are coming from China in ever larger dimensions. Models are coming from the USA that are being made openly available more often than before. And new projects are announcing themselves in Europe: there is news from Switzerland and, after a longer wait, from Germany too.
For companies that build AI functions into websites, portals or internal tools, this is not a side note. Because the question of which models are openly available helps determine where these functions can be operated — and therefore also which data leave your own network.
What "open" means in this context
Openly available generally means: the model weights, i.e. the trained parameters, are available for download and may be run on your own infrastructure. This is not the same as open source in the classic sense, because training data and training code are usually not disclosed and the licences vary in strictness.
In practice, however, the operational aspect is what matters most. A model whose weights you can load runs on your own server, in your own cloud instance or in the data centre of a European provider. A model that is only reachable as a service via an interface does not.
Why size is an operational question
When models grow, the demands on memory and computing power grow with them. Very large models are difficult to operate yourself even when the weights are freely accessible: they need several powerful accelerator cards, corresponding RAM and an operations team that has this under control.
That is precisely why smaller models are often the more interesting news for mid-sized companies. They can be run on manageable hardware, respond faster and cause calculable costs in continuous operation. The price for this is usually less general knowledge and less performance on complex tasks.
The practical benefit of small models
Many tasks in web and software projects are narrowly defined. For such cases, a smaller model is often sufficient:
- Classifying and tagging: assigning enquiries from forms to a category, automatically providing editorial content with keywords.
- Summarising: shortening long documents or ticket histories to the essentials.
- Rewriting and checking: generating text variants for teasers, proofreading against style guidelines.
- Structuring: extracting clean data fields from unformatted texts.
- Semantic search: translating content into vectors for a search function that map semantic similarity.
For tasks with high demands on reasoning or depth of expertise, large models remain at an advantage. This argues for a mixed architecture in which simple enquiries are answered locally and only selected cases go to an external service.
Sovereignty is more than data protection
Data protection is the obvious argument for operating in-house: if personal data or trade secrets are not transferred to an external provider, a whole block of checks, contracts and risk assessments falls away.
Technical continuity is just as important. A model that is local does not disappear at the end of the quarter, does not change its behaviour without prior notice and does not become more expensive overnight. Anyone who builds productive functions on a model is buying, above all, predictability with openness.
On the other hand there is the effort: operation, updating, monitoring and securing then lie within your own area of responsibility. This calculation turns out differently depending on the company and the use case.
What this means for the architecture
The current movement in the market cannot be absorbed by a one-off selection decision. Anyone who commits to a model today will have to make this decision again within a manageable time. What makes sense, therefore, is an intermediate layer in your own application that keeps the model interchangeable.
Concretely, this means: prompts, post-processing and error handling lie in your own code, not in a provider-specific library. For every AI function there is a small set of test cases with expected results against which a model change can be evaluated. And there is a defined path for what happens when the model does not respond or delivers unusable output.
In a TYPO3 project this can be well encapsulated: the AI connection sits in its own service, and editors continue to work with their usual fields and masks. Whether a local model or an external interface responds in the background need not concern anyone in the backend.
Assessment
The variety of openly available models increases the scope for deciding where AI functions are operated — and precisely this scope is the real gain for data-protection-sensitive applications. For projects, it is therefore worth investing less energy in choosing the supposedly best model and more in an architecture that allows switching without rebuilding. Anyone who plans this way can test smaller European models today without blocking the path to larger systems.
Sources
- Model-Schau: Offene Modelle aus den USA, große aus China und kleine aus Europa — heise developer News