On July 24, 2026, ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5, a multimodal upgrade to Switzerland's fully open, state-backed language model. The flagship model — plus a new "Apertus Mini" suite of 16 compact, distilled versions — now understands images and audio alongside text, with improved reasoning and tool use. Everything, including training data and code, ships under an Apache 2.0 license. As ETH Zurich's Imanol Schlag put it, the project is "not about competing with frontier models from private companies, but about providing trustworthy, transparent AI."
That framing matters. Apertus is explicitly positioned, per the CSCS release, as building "a sovereign AI infrastructure that provides an open alternative to proprietary commercial models" — a phrase Bern has leaned on since the original model shipped in September 2025. The model already has real deployments: translation tooling for Ticino's cantonal government, the MeditronFO medical-AI framework at EPFL, and workflow tools at the Basel newsroom Bajour, whose editor Samuel Hufschmid noted the appeal is that "we know what we're building on." Trained on the Alps supercomputer in Lugano, Apertus is genuinely one of the largest fully open models in the world — full weights, training corpus provenance, and methodology are public in a way no US or Chinese frontier lab offers.
The steelman for sovereign AI
The case for building this isn't abstract nationalism. Switzerland's own defense establishment has formally raised concerns about Microsoft dependency reaching the federal government — a legitimate continuity-of-government concern, not a talking point. A government that cannot audit, modify, or run critical AI tooling without a foreign vendor's continued cooperation has a real exposure, whether that vendor is well-intentioned or not. Public-interest institutions like hospitals and courts also have a defensible reason to prefer a model whose training data and weights they can actually inspect, rather than trusting a black box's safety claims. On that narrower point — transparency and auditability as public goods, independent of raw model quality — Apertus is a legitimately useful piece of infrastructure, and public funding for it is well justified.
Where the sovereignty claim overreaches
But "sovereign" is doing more work in the messaging than the facts support. Apertus trained on Nvidia GPUs inside Alps, a machine that itself cost roughly CHF 100 million to build — a fraction of what EPFL's Marcel Salathé says would be needed for real independence, and ETH's Alexander Ilic estimates Switzerland will need 20 to 100 times more compute within a decade just to stay relevant. Salathé is blunt about the ceiling: full sovereignty is "completely impossible for a small country like Switzerland," given globally concentrated control over chips, power, and cloud infrastructure. The Swiss AI Initiative itself runs on roughly CHF 20 million through 2028 — against Singapore's $1.6 billion AI compute commitment, the gap is not close.
The sharper contradiction sits inside the federal government itself. The same month Apertus first shipped, Bern renewed its Microsoft Enterprise Agreement for CHF 140 million, running 2025 through 2027 — and, per SRF reporting relayed by SWI swissinfo, federal and cantonal bodies have spent over CHF 1.1 billion on Microsoft licensing in the past decade, with roughly 54,000 administration workstations now running Microsoft 365. The Federal Chancellery says it wants to reduce that dependency "step by step and in the long term," pointing to Schleswig-Holstein's open-source migration in Germany as a model — but no deadline has been set, and the licensing renewal itself was reportedly awarded without competitive tender.
The proportionate read
None of this makes Apertus a bad investment — it's a genuinely good one, on transparency-and-auditability grounds alone, and the Apache 2.0 release is a real public good other governments and researchers can build on. The mistake is letting a research-model release stand in for an infrastructure strategy. Publishing an open LLM doesn't reduce a chip dependency measured in GPUs, and it doesn't touch the CHF 1.1 billion in day-to-day software lock-in that actually determines how exposed the Swiss state is to a foreign vendor's decisions. If Bern wants sovereignty claims to hold up, the higher-leverage move is fixing procurement — competitive tenders instead of direct awards — and funding the multi-decade compute buildout Ilic says is coming, not treating a model release as the finish line. Switzerland should keep building Apertus. It should stop calling it sovereignty until the procurement and compute numbers actually point the same direction.