ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5 on July 24, 2026 — the second major iteration of Switzerland's fully open large language model, trained on the Alps supercomputer in Lugano. The update adds multimodal understanding of images and audio, stronger reasoning and tool use, and a companion suite of 16 lightweight "Apertus Mini" models for constrained hardware. It is released, like its predecessor, under the permissive Apache 2.0 license, meaning the weights, training data and code are all public.
The more consequential story isn't the model card. It's that a Swiss canton is now running government paperwork through it. In March 2026, the Information Systems Centre (CSI) of Canton Ticino deployed an in-house translation tool built on a fine-tuned Apertus-8B, developed by Lugano-based Artificialy. The fine-tuned version scores 94% on internal benchmarks, versus 90% for the stock 8B model and 92% for the larger 70B — evidence that a small, sovereign-hosted model can out-perform a bigger general-purpose one on a narrow task. About 100 cantonal employees who need daily translation now route sensitive documents through GPUs sitting in the canton's own data center, never touching a commercial API.
The steelman for caution
There's a real case for treating this rollout skeptically rather than celebrating it. A 94% benchmark score on internal tests is not the same as 94% reliability on the specific legal, medical or administrative documents a canton actually handles — and a mistranslated clause in an official document carries consequences a chatbot typo doesn't. Switzerland also has no AI-specific statute yet: the Federal Council's own regulatory roadmap says a consultation bill implementing the Council of Europe's AI Convention, built around transparency, data protection and non-discrimination, won't be ready until the end of 2026. Ticino is running government workloads on a foundation model ahead of the rules that will eventually govern it. And the Federal Data Protection and Information Commissioner (FDPIC) has been explicit that Switzerland's existing Federal Act on Data Protection already applies in full to AI-supported processing — high-risk uses need a data protection impact assessment, and providers must disclose an AI system's purpose, functionality and data sources. Regulators watching this space are not being alarmist; they're pointing at a real gap between deployment speed and oversight infrastructure.
Why open-and-local still wins the trade
That caution argues for scrutiny of this particular deployment, not against the architecture. The entire reason Ticino avoided that data protection impact assessment headache is that the model runs on-premises: sensitive government documents never leave cantonal infrastructure, unlike a commercial cloud translation API where data sovereignty is a contractual promise rather than a physical fact. Apertus's developers say they built the model with Swiss data protection and copyright law, and the EU AI Act's transparency obligations, in mind from the training-data stage — filtering for opt-outs and personal data before training rather than bolting on compliance after the fact. That's a genuinely better starting posture for a government user than adopting a black-box commercial model and hoping vendor terms hold up under FDPIC scrutiny later. And because the weights and training data are public, Ticino's 94%-versus-90% benchmark claim can actually be audited by outside researchers — a check that's structurally impossible with closed frontier models.
The funding problem is the real story
The more interesting tension isn't sovereignty versus openness — it's ambition versus budget. Apertus has roughly CHF 20 million committed through 2028, and the Alps supercomputer represents about CHF 100 million of investment. ETH Zurich's Alexander Ilic has said the initiative will need "20 to 100 times more computing power" within a decade to stay relevant, while researchers quoted in Swiss press note operations alone are short by roughly CHF 10 million a year. Compare that to Singapore's CHF 1.3 billion national AI push, or the multi-billion-euro compute buildouts underway across the EU and US. Switzerland is trying to prove a small, transparent, publicly accountable model can compete on trust and auditability rather than raw scale — and Ticino's translation tool is the first real evidence that proposition can work in production, not just in a benchmark paper. But a sovereignty strategy that can't fund its own compute roadmap is a demonstration project, not an infrastructure plan.
The right regulatory response to Ticino's rollout is not to slow it down while Bern finishes its AI Convention consultation — it's to let cantons keep adopting open, auditable, locally-hosted models under the data protection law that already binds them, while treating the funding gap Ilic and others describe as the actual policy failure worth fixing. An open sovereign model that Switzerland can't afford to scale is a worse outcome for both privacy and innovation than one it invests in properly.