On July 24, 2026, ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5, the second generation of Switzerland's fully open large language model. Built under the Swiss AI Initiative and trained on the Alps supercomputer in Lugano, the new version adds image and audio understanding, a four-fold expansion of its context window to 262,144 tokens, and a companion suite of 16 distilled "Apertus Mini" models for resource-constrained deployments. Released under the Apache 2.0 license, the model's weights, training data, and code are all public — a deliberate contrast to the closed frontier labs. As ETH's Imanol Schlag put it, Apertus "is not about competing with frontier models from private companies, but about providing trustworthy, transparent AI that organizations can build upon with confidence."
That framing would be easy to dismiss as academic posturing if not for Ticino. The southern canton's Information Systems Centre began piloting a fine-tuned Apertus-8B translation tool built by the Lugano firm Artificialy, and roughly 100 cantonal employees now use it to translate sensitive government documents across Italian, German, French, Spanish, Romanian, and Ukrainian — entirely on cantonal infrastructure, with no document ever leaving Swiss servers. The fine-tuned model scored 94% accuracy in benchmarking, edging out both the base Apertus-8B (90%) and Meta's Llama Scout (91.4%). This is the detail that matters for policy: Ticino didn't adopt Apertus out of loyalty to a national project. It ran a bake-off, and the sovereign option won on merit as well as on data residency.
The Case for Sovereign Infrastructure
There is a serious argument behind Switzerland's AI sovereignty push, and it deserves to be stated plainly before it's contested. Government translation work routinely touches asylum files, health records, and law-enforcement material — exactly the category of data where sending text to a foreign-hosted commercial API creates real exposure, whether to foreign subpoena law, vendor breach, or simple loss of control over retention. Switzerland's Federal Chancellery has separately acknowledged that the federal administration is "factually dependent on Microsoft's Office products," and that this dependency "gives the company a strong negotiating position regarding prices, which has corresponding cost consequences." A national government maintaining the capability to run its own models, on its own hardware, under its own legal jurisdiction, is a reasonable hedge against vendor lock-in and geopolitical risk — not merely symbolism.
The Unfinished Half of the Bet
But Ticino's translation tool is one workflow in one canton. The federal government that funds the Swiss AI Initiative is, at the same time, deeply entangled with the vendor Apertus is implicitly positioned against. In late 2024, the Federal Office for Buildings and Logistics renewed the government's Microsoft Office 365 licensing for CHF 140 million without competitive tender, on the grounds that the service remained "weiterhin zwingend erforderlich" — still absolutely necessary. Professor Matthias Stürmer, who studies administrative digitalization, called it "a fat Christmas present from us taxpayers," awarded without a tender process open-source vendors could even compete for. The Federal Chancellery's own language — "factually dependent" — is not the description of a government that has kept its options open.
This is the gap pro-innovation observers should press on: building an excellent open model is necessary but not sufficient for sovereignty. Sovereignty is demonstrated in procurement decisions, not press releases. Ticino's rollout is proof that the substitution is technically viable at least for narrow, well-scoped tasks; it is not yet evidence that Bern intends to replace anything load-bearing. A national AI model that federal agencies don't actually use for federal work functions as reputational insurance, not infrastructure.
Regulation Is the Part Switzerland Is Getting Right
Where Switzerland's approach holds up well is regulatory restraint. Rather than legislate a comprehensive AI Act along EU lines, the government has opted to ratify the Council of Europe's AI Convention and adapt existing sectoral law — with binding and non-binding measures for sensitive domains like healthcare and autonomous transport due by the end of 2026. That is the proportionate path: it lets a project like Apertus prove itself through adoption, benchmarking, and canton-level experimentation rather than forcing every deployment through a compliance gate calibrated for hypothetical worst cases. The EU's risk-tiered AI Act imposes conformity assessments on systems well short of Ticino's narrow translation use case; Switzerland's lighter model let a 100-employee pilot ship in months, with the accuracy numbers doing the persuading instead of a regulator's checklist.
The test now is whether that light-touch regulatory model gets paired with a genuine procurement shift. Switzerland has shown, in Ticino, that open-weight sovereign AI can out-benchmark commercial alternatives on a real government workload. The harder proof point — a federal agency actually migrating a Microsoft-dependent function to Apertus, on the same competitive terms Ticino applied — hasn't happened yet. Until it does, Apertus 1.5 is a genuinely impressive technical achievement operating alongside, rather than inside, the government's actual infrastructure.