In May 2026, Lausanne University Hospital (CHUV) began running Meditron — an open-source medical large language model developed at EPFL — alongside physicians in its emergency department. It is a modest pilot in one hospital. It is also the clearest real-world test yet of a national strategy that most policymakers elsewhere have treated as mutually exclusive: building sovereign AI capacity with public money and refusing to wrap it in a comprehensive, EU-style regulatory framework.
What Switzerland actually built
Meditron is being rebuilt on Apertus, the fully open foundation model that EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) released on 2 September 2025 under a permissive Apache-2.0-style licence. Apertus is unusual: its architecture, weights, training data and methods are all published, it ships in 8-billion and 70-billion-parameter sizes, and it was trained on 15 trillion tokens across more than 1,000 languages — 40% of it non-English, including Swiss German and Romansh (ETH Zurich; EPFL).
An Apertus-based Meditron is, in principle, fully sovereign: Swiss-designed architecture, a Swiss-trained foundation model, Swiss-curated medical data, deployed on hospital servers. That last point matters most in healthcare. As an open model running on CHUV's own infrastructure, Meditron can be fine-tuned on the hospital's patient records, clinical notes and outcomes without any of that data ever leaving the building — and without sending sensitive records to a US cloud provider. More than 300 clinicians evaluated the model on hypothetical cases throughout 2025 before this live trial began (zuerich.ai).
The whole effort sits under the Swiss AI Initiative, financed by the federal government to the tune of CHF 20 million through 2028 (SWI swissinfo). That is a rounding error next to the tens of billions flowing into private frontier labs — and that frugality is precisely the point.
The regulatory bet
On 12 February 2025, the Federal Council settled Switzerland's regulatory course. It will ratify the Council of Europe's Framework Convention on Artificial Intelligence — the first binding international AI treaty — and amend domestic law accordingly, with a public consultation draft due by the end of 2026 (admin.ch). Crucially, the Convention will apply mainly to state actors. Where new rules are needed for the private sector, they are to be as sector-specific as possible, with cross-cutting obligations confined to fundamental-rights areas such as data protection. There will be no Swiss AI Act.
The case for the opposite choice deserves a fair hearing. A medical LLM advising emergency physicians is close to a textbook high-risk system: errors can kill, the pressure is acute, and the EU AI Act would subject exactly this use case to conformity assessments, risk-management systems, human-oversight mandates and post-market monitoring. Proponents of horizontal regulation argue that a single, predictable rulebook spares companies from reconciling a patchwork of sectoral regimes and gives the public a clear floor of protection. That is a serious argument, and the harms it guards against are real.
But Switzerland's wager is that the floor already exists — in medical-device law, professional-liability rules, the revised Federal Act on Data Protection, and now a ratified international treaty — and that a second, AI-specific horizontal layer would mostly add cost and delay. The Meditron pilot is the test of that proposition. Nothing about deploying an AI tool in a Swiss ER was blocked for want of a dedicated AI statute; the existing duties of clinical oversight, data protection and device safety apply on their own terms. A team could move from a 2025 evaluation phase to a 2026 live deployment because the regulatory question was "does this comply with the rules that already govern emergency medicine and patient data," not "does this clear a new conformity-assessment pipeline."
Why the two halves fit together
The deeper insight is that public investment in open infrastructure and light-touch regulation are complements, not substitutes. Heavy horizontal rules tend to favour incumbents who can absorb compliance overhead — the very large foreign labs Switzerland's sovereignty push is meant to reduce dependence on. By funding an open model instead and keeping the rulebook lean, Switzerland lowers the barrier for its own hospitals, universities and startups to build on domestic capacity. Sovereignty here is achieved through openness and access, not through a protective wall of mandates.
There are real caveats. Open weights do not by themselves guarantee clinical safety; a model fine-tuned on local data can still hallucinate a dose or miss a contraindication, which is why physician oversight remains the binding control. CHF 20 million will not keep pace with frontier compute indefinitely, and "sovereign" capacity that lags badly on quality helps no patient. The end-2026 consultation could still drift toward heavier cross-sector obligations under EU pressure, blunting the model's advantage. And a single ER pilot is not yet evidence of system-wide success.
Still, the direction is instructive. While much of the policy debate frames AI governance as a choice between innovation and protection, Switzerland is quietly running the experiment that says you can have proportionate rules and a credible public alternative to closed foreign models — and that the two reinforce each other. If Meditron proves clinically useful in Lausanne's emergency room, the lesson other small and mid-sized democracies should draw is not "copy the AI Act." It is: invest in open infrastructure, regulate the use rather than the technology, and trust the laws you already have.