Switzerland Built Its Own LLM — and Education Is a Named Target Domain
On 2 September 2025, EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS) released Apertus, Switzerland’s first large-scale open language model. It was trained on the Alps supercomputer in Lugano on roughly 15 trillion tokens across more than 1’000 languages, including Swiss German and Romansh. It ships in 8-billion and 70-billion parameter versions under an Apache 2.0 license that permits use in education, research, and commercial projects.
The detail that matters for education leaders sits in the roadmap. The project team has named education, alongside law, health, and climate, as a priority domain for future specialized versions. Switzerland now has a public, fully documented model that institutions can run on their own infrastructure, inspect end to end, and adapt to curricula and administrative workflows.
That release did not happen in isolation. It is 1 of at least 4 signals from the past 12 months that change how Swiss schools, universities, and training providers must plan their AI deployments.
The Signal: Regulators Are Closing the Door on Casual Cloud AI in Schools
Swiss education institutions process some of the most sensitive personal data that exists: learning records, psychological assessments, health notes, family circumstances of minors. Under revDSG and cantonal data protection law, much of this qualifies as besonders schützenswerte Personendaten — data requiring special protection.
Here is what changed:
- privatim, the conference of Swiss cantonal data protection commissioners, adopted a resolution on international cloud solutions that explicitly covers education institutions. It states that outsourcing sensitive personal data to international providers is inadmissible in most cases.
- Educa, the specialist agency of the EDK and SBFI, published its report on data usage policy for the Swiss education space (EDK decision of 26 June 2025). Its recommendation is direct: data of high sensitivity may only be processed on Swiss infrastructure. International tools remain possible for less critical applications, but only after thorough review and with clear contractual agreements.
- The Federal Council’s report on Swiss digital sovereignty names education as a covered domain, tying digital skills and infrastructure decisions in schools to national sovereignty goals.
- The Federal Council confirmed its sector-specific AI regulation approach on 12 February 2025 and instructed the FDJP to prepare a consultation draft implementing the Council of Europe’s AI Convention by the end of 2026. The convention applies primarily to state actors — and public schools and universities are state actors.
Read together, these signals define a corridor. AI in Swiss education is welcome, even encouraged. AI that routes student data through international consumer platforms is on a path to becoming legally indefensible.
Why This Creates a Concrete Problem for Education Institutions
Picture the current reality in a typical Swiss Gymnasium or Fachhochschule in mid-2026. Teachers use ChatGPT or similar tools to prepare lessons, sometimes pasting in student work for feedback. Administrators draft parent communications with consumer AI tools. Students submit AI-assisted homework. Almost none of this runs on Swiss infrastructure. Almost none of it went through the application checklist that Educa recommends in its «Datenschutzkonforme Schule» dossier.
Each of these individual actions looks harmless. Together, they create 3 observable problems:
- Uncontrolled data flows. Student names, grades, and behavioral notes end up in prompts sent to servers outside Switzerland, with no processing record. Under the privatim resolution, a school leadership that tolerates this is tolerating a practice the cantonal supervisory authorities have already classified as mostly inadmissible.
- No auditability. Commercial frontier models disclose neither training data nor full processing chains. When a parent, a cantonal data protection officer, or a journalist asks how an AI-supported decision about a student was made, the institution cannot answer.
- A widening gap between policy and practice. The EDK reviewed its digitalisation strategy in March 2026. Cantons are building data governance structures. Meanwhile, actual AI usage in classrooms runs ahead of every guideline. The institutions that close this gap deliberately will be in a defensible position when the AI Convention implementation arrives. The ones that don’t will be retrofitting compliance under time pressure.
The same logic applies to private education groups, edtech companies, and corporate training providers selling into the Swiss market. Their buyers — school boards, cantonal education departments, university procurement — are being told by Educa and privatim to demand Swiss infrastructure and documented data handling. Vendors who cannot answer those questions are already losing tenders.
5 Options for Education Institutions and Providers Acting in 2026
1. Run Open Models Like Apertus on Controlled Infrastructure
Apertus can be downloaded from Hugging Face or accessed via Swisscom’s sovereign AI platform. Because architecture, weights, and training data documentation are fully published, an institution can verify what the model contains and run it on-premise or in a Swiss data center. This removes the international data transfer problem at the root. The trade-off is real: a 70B open model requires serious compute, and raw model quality trails the largest commercial systems. For education use cases — lesson support, administrative drafting, multilingual communication in national languages — the gap is often acceptable, and the compliance position is fundamentally stronger.
2. Move Sensitive Workloads to Swiss or On-Premise Deployments
For institutions that need higher capability than open models currently deliver, the deciding factor is where processing happens and what leaves the building. A custom AI system built from an institution’s own documents — regulations, curricula, internal FAQs, administrative procedures — and deployed on Swiss or on-premise infrastructure satisfies the Educa recommendation that high-sensitivity data stays on Swiss infrastructure. This is the deployment model Lab51 builds for organizations in regulated sectors: fully custom systems from the client’s own data, running where the client’s compliance rules require.
3. Implement the Educa Data Governance Toolkit Before Buying Anything
Educa’s «Datenschutzkonforme Schule» dossier provides 2 concrete instruments: a processing record (Bearbeitungsverzeichnis) and an application checklist. Institutions that complete both gain something most currently lack: an inventory of what data exists, where it flows, and which applications touch it. This costs staff time, not license fees, and it is the prerequisite for every serious AI procurement decision. A vendor evaluation without a processing record is guesswork.
4. Build an AI System Inventory Ahead of the 2026 Consultation Draft
The FDJP’s consultation draft on AI regulation is due by the end of 2026, with transparency, data protection, non-discrimination, and oversight as its core areas. Legal practitioners already recommend that organizations maintain an inventory of every AI system in use: purpose, risk category, data involved, governance measures. Public education institutions, as state actors, sit squarely in the AI Convention’s scope. Starting the inventory now converts a future compliance scramble into a routine documentation exercise.
5. Invest in AI Literacy for Staff, Not Just Tools
Educa’s own report stresses that technical measures alone are insufficient. Schools in and around Zurich are already restructuring programs around AI literacy — critical evaluation, data awareness, ethics — across subjects. The European Commission updated its ethics guidelines for teachers in March 2026 under the Digital Education Action Plan, giving institutions a ready framework. Teacher training determines whether governance rules are followed in practice or bypassed with private accounts. Every deployment plan that skips this line item fails at the classroom door.
Why Now: The Window Between Guidance and Law
Assumption stated openly: exact enforcement timing is not yet known, since the AI Convention bill goes to consultation at the end of 2026 and ratification requires parliamentary approval. What is known is the direction. privatim has ruled on international clouds. Educa has published the reference framework cantons will use. Apertus exists as proof that sovereign, education-suitable models are feasible. And the EDK closed its digitalisation strategy review in March 2026, which means the next planning cycle is being written now.
Institutions and providers that align their AI architecture with Swiss infrastructure requirements in 2026 do it on their own schedule and budget. Those that wait will do it under a statutory deadline, with every competitor and canton making the same requests to the same few implementation partners at the same time.
Switzerland has chosen its path on AI in education: sovereign infrastructure, documented data flows, sector-specific rules, and a public model built for exactly this purpose. The decisions are no longer conceptual. They are procurement decisions, and the institutions making them deliberately this year will set the standard the rest will be measured against.
