ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5 on July 24, 2026, the first multimodal update to Switzerland's fully open large language model. The 8B-parameter model was continued-pretrained on 4 trillion additional tokens of text, image, and audio data, the 70B model on 2 trillion, and the release ships alongside a new "Apertus Mini" suite of 16 compact models for lighter deployment, according to ETH Zurich's AI Center. Everything — weights, training data, and methodology — remains published under an Apache 2.0 license, as it was for the original Apertus release in September 2025.
That first release, trained on the Alps supercomputer in Lugano using over 10 million GPU hours from CSCS and roughly 40% non-English data (including Swiss German and Romansh), was explicitly framed by its builders as more than a research artifact. EPFL's Martin Jaggi called it "a blueprint for how a trustworthy, sovereign, and inclusive AI model can be developed," per ETH Zurich's original press release. Apertus 1.5 extends that pitch into multimodal territory just as Switzerland is deciding, in parallel, how much its government should actually depend on foreign AI infrastructure.
The case for a public model
The steelman for a state-adjacent, fully open model is stronger than "innovation theater." Canton Ticino already runs Apertus in-house to translate sensitive government documents, and the Basel newsroom Bajour deploys a locally-hosted version for political reporting — precisely because neither wants classified or editorially sensitive material transiting a US vendor's API, per EPFL's release notes. A model whose training data and weights are fully auditable lets a hospital, court, or newsroom verify what the system was actually trained on, rather than trusting a black-box provider's safety card. For a small, multilingual country wary of concentrating dependency in two or three American labs, that auditability is a legitimate public good, not just symbolism — and open weights mean Switzerland isn't locked into a single vendor's pricing or shutdown risk down the line.
Where the sovereignty story gets complicated
But Apertus's existence doesn't mean Swiss digital sovereignty policy has caught up with the model. The Federal Chancellery's own AI strategy, last updated for 2026, centers on building AI competence and "trustworthy use" inside the administration — it does not mandate Apertus, or any open model, for federal workloads. The Federal Council's actual regulatory vehicle is a horizontal AI bill implementing the Council of Europe's AI Convention, due for consultation by the end of 2026, covering transparency, data protection, and supervision — a procedural track that runs independently of which models any ministry actually buys.
Meanwhile the federal government is renewing its Microsoft licensing at a cost of roughly CHF 140 million over two years, according to SWI swissinfo.ch's 2026 AI policy roundup — the same reporting that flags "digital sovereignty" as a stated 2026 priority aimed at reducing dependence on individual, often American, suppliers. That's not hypocrisy so much as the ordinary lag between a research achievement and a procurement bureaucracy; Microsoft 365 and Azure are deeply embedded in federal IT, and ripping them out for an LLM's sake would be its own kind of recklessness. But it does mean Apertus's "sovereign alternative" framing is currently aspirational for most of the Swiss state, not descriptive of how Bern actually runs.
Why the open-weight approach is still the right call
The useful reading of Apertus 1.5 isn't "Switzerland has solved AI sovereignty" — it's that Switzerland has built genuine optionality. A hospital testing Meditron (the Apertus-derived medical model CHUV began piloting in emergency rooms in May 2026, per the swissinfo report) or a canton translating government filings gets a real, inspectable alternative to sending data to a proprietary API, without needing new legislation to get there. That's a better model for middle powers than either uncritical dependence on US frontier labs or a defensive, protectionist mandate forcing agencies onto homegrown tools before they're ready. Open weights let adoption happen use-case by use-case, driven by actual trust and cost tradeoffs rather than top-down procurement rules — which is exactly the kind of proportionate, market-led path this publication has argued for elsewhere in AI governance.
The risk is conflating the two tracks in public messaging. Every Apertus press release invokes "sovereign AI" while the concrete Swiss AI Convention bill working through Bern says nothing about which vendors agencies must use. If Switzerland wants Apertus to be more than a well-funded academic proof of concept, the honest next step is a federal procurement preference — piloted, not mandated — for auditable models in the sensitive-data use cases (health, courts, immigration) where the sovereignty argument is actually strongest. Absent that, Apertus 1.5 is a genuinely impressive research release riding on a policy narrative the government hasn't yet funded.