Self-hosted and offline AI models in schools

Most well-known AI chat programs run as an internet service: inputs are sent to the provider's data centres, processed there and the answer sent back. There are also language models whose weights are publicly available and which can be run on your own hardware. Such open models come in many sizes, from small models that run on a laptop to large ones that need servers with several graphics processors.

“Self-hosted” can mean different things: a computer in the school, a data centre of the school authority or state, or a service provider running the models exclusively for the school in a contractually defined environment. “Offline” means the model works without an internet connection — on devices in the classroom, for instance.

Self-hosting solves several problems schools have with cloud services.

Self-hosted models are usually smaller than the largest cloud models. For many school tasks — explaining, practising, summarising, asking questions — that is enough, especially when the model is tied to checked material. With complex reasoning, specialised knowledge or long texts, smaller models are weaker and make more mistakes.

Then there is operation: hardware with powerful graphics processors is expensive and needs power and cooling. Models must be updated, security gaps closed, outages fixed. For a single school this rarely makes sense; for a school authority, a state or a specialised provider serving many schools, it is more likely to.

Self-hosting does not settle every question. A self-hosted model still needs rules: What may it answer, how is it tied to curricula, how are children protected from unsuitable content, who sees the logs? Data protection law continues to apply when personal data are processed; the AI Act applies regardless of where the model runs.

Guardrails before and after every model call are also important: fixed rules the model cannot override, a check of every input for attempts to circumvent those rules, and a check of every output. These guardrails are needed wherever the model runs.

Cloud, self-hosting or a mix — the decision depends on a few questions. Which tasks should AI take on, and how capable must it be for them? How sensitive are the data? Which specialists are available for operation? Are there state solutions or framework agreements? And what does the total cost look like over several years, including staff?

A staged approach often makes sense: central, self-hosted models for everyday use with pupil data, supplemented by clearly governed services for special tasks involving no personal data.

For the planned Universe Classroom offer, the models are intended to run in Universe's own operation, not with a third-party provider. What this operating model looks like in detail is still being clarified. Curriculum fidelity via teacher-confirmed learning objectives, control by teachers and protection of minors are part of the plan.

The current platform already follows these rules: data and processing are in the EU, as a hard rule in the code; before every model call there are immutable rules, every input is checked deterministically and every output too, with a log. The platform's avatar engine is also self-operated, on rented graphics servers.