On July 24, 2026, ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5, the second major iteration of Switzerland's fully open large language model. The new version adds multimodality — it now processes images and audio alongside text — and ships alongside Apertus Mini, a family of 16 compact distilled models built for phones, laptops and other resource-constrained hardware. Both are released under the Apache 2.0 license, meaning anyone, including commercial competitors, can inspect, modify and redeploy them freely.
That openness is the point. Apertus emerged from the Swiss AI Initiative, launched in December 2023 as a joint venture of ETH Zurich, EPFL and CSCS, drawing on more than 800 researchers across ten-plus Swiss institutions. The original model, released in September 2025, was trained on 15 trillion tokens spanning over 1,000 languages — including Swiss German and Romansh — with 40% non-English data, and shipped in 8-billion and 70-billion parameter sizes. Unlike Meta's Llama or Mistral's models, which release weights but keep training data and filtering methods proprietary, Apertus publishes its entire pipeline. EPFL's Martin Jaggi called it a "blueprint for how a trustworthy, sovereign, and inclusive AI model can be developed."
The sovereignty argument, and its limits
The project sits inside a broader European anxiety about dependence on US and Chinese AI infrastructure. Nvidia chips power roughly 90% of the world's AI models, Apertus included, and Switzerland has no domestic chip fabrication to fall back on. Federal spending on the effort totals roughly CHF 100 million for the Alps supercomputer plus a further CHF 20 million through 2028 for the Swiss AI Initiative itself — a real commitment, but one dwarfed by Singapore's roughly $1.6 billion sovereign AI investment, in an economy about a third the size of Switzerland's. Researchers quoted by SWI swissinfo.ch argue Switzerland needs 20 to 100 times more compute over the next decade to stay relevant, and warn that maintenance and energy costs alone will run CHF 10 million annually going forward.
Apertus has also drawn genuine criticism, not just skepticism from onlookers. Its 70-billion parameter model trails GPT-4-class systems on standard reasoning and coding benchmarks, and some testers reported Swiss German and Romansh outputs that were barely usable. That criticism deserves a fair hearing: a publicly funded model that can't do the job commercial alternatives already do isn't obviously a good use of taxpayer money, and "it's transparent" is a weaker argument if the transparency doesn't translate into anything a Swiss hospital, ministry or bank can actually deploy. EPFL's Philippe Cudré-Mauroux has pushed back that Apertus was never meant to be a ChatGPT competitor — it's foundational infrastructure meant to be fine-tuned for specific institutional tasks, where auditability of training data matters more than leaderboard scores. That's a legitimate distinction, but it also means Apertus's real test is adoption by the public administrations and companies it's built for, not benchmark parity with frontier labs.
Funding openness while keeping regulation light
What makes the Apertus story more than a research footnote is that it's happening alongside a deliberate regulatory choice. On February 12, 2025, the Federal Council decided against a comprehensive AI-specific statute on the EU model. Instead, Switzerland will ratify the Council of Europe's Framework Convention on AI and adapt existing sector-specific law — data protection, financial-markets rules, anti-discrimination law — rather than layering a new horizontal AI Act on top. The Federal Department of Justice and Police is due to publish a consultation draft by the end of 2026, focused narrowly on transparency, data protection, non-discrimination and oversight.
The steelman for the EU's approach is real: the AI Act's risk-tiered obligations, provider registration and conformity assessments exist because generative models are already making consequential decisions in hiring, credit and healthcare with little external visibility, and voluntary codes have a mixed record of actually constraining powerful incumbents. A model as opaque as most frontier commercial systems arguably needs exactly the kind of mandated disclosure the AI Act imposes.
But Switzerland's bet is that it can get much of that accountability for free by funding transparency at the source rather than mandating it after the fact. A model whose training data, filtering choices and architecture are public by design, as Apertus's are, gives regulators, researchers and competitors more to audit than a compliance filing ever would — without freezing a sector-specific rulebook around today's technology. Pairing that with narrow, sector-anchored rules rather than a comprehensive AI Act lets Switzerland avoid locking in compliance costs before anyone knows which AI harms actually materialize. For a small, open economy that can't outspend Washington or Beijing on compute, publishing the model instead of just regulating around it is the more defensible way to build public trust in AI — provided institutions actually adopt Apertus rather than treating it as a prestige project that ships benchmarks nobody uses.