A public model grows up
On July 24, 2026, ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5, adding image and audio understanding to what began in September 2025 as a text-only open-weight model. The upgrade also brought Apertus Mini, a 16-model family distilled and quantized for phones and edge devices — a signal that Switzerland wants Apertus used, not just admired.
The original Apertus was trained on 15 trillion tokens across more than 1,000 languages, including Swiss German and Romansh, in 8B and 70B parameter sizes, on the Alps supercomputer's roughly 10,752 NVIDIA GH200/Grace-Hopper GPUs at CSCS in Lugano. Both generations ship under Apache 2.0 — full weights, training code and data documentation, unrestricted commercial use. Downloads passed one million within a year.
The case for building your own model
The strongest argument for a public LLM isn't nostalgia for Swiss neutrality — it's dependency. Every government running services on a foreign frontier model is one export licence, pricing change or policy dispute away from disruption. A model trained "with due consideration to Swiss data protection laws, Swiss copyright laws, and transparency obligations" gives Bern something no vendor contract can: legal clarity over what data went in and who is accountable for what comes out. Full weight and data transparency also lets outside researchers actually audit the thing, unlike the closed frontier labs it's positioned against. Switzerland has also put this to work domestically: the Canton of Ticino uses Apertus for official-document translation, and Swisscom deployed it on a "sovereign Swiss AI Platform" for business customers.
Where the sovereignty framing strains
That case is real, but Switzerland's own numbers undercut the more sweeping version of it. The federal government has committed CHF 20 million to the Swiss AI Initiative through 2028 — funding a genuinely useful research asset — while the federal administration separately renews Microsoft licensing for roughly CHF 140 million over two years. As ZHAW's Thilo Stadelmann put it: "On the one hand we talk about technological sovereignty, on the other hand we continue to rely on a few large foreign players." Owning a model's weights is not the same as owning the compute, cloud, and OS layers most Swiss institutions still run on. Apertus is a genuine achievement in transparency; it is not, by itself, sovereignty in any operationally meaningful sense, and officials should stop implying otherwise.
There's a second honest caveat: independent assessments still place Apertus behind frontier proprietary models on raw capability. That's an acceptable trade for a publicly documented, auditable model — but it means Apertus should be marketed as accountable infrastructure, not as competing head-to-head with the largest commercial systems.
Why this is the right complement to a light regulatory touch
Where Switzerland gets the sequencing right is in refusing to pair this infrastructure push with an EU-style horizontal AI Act. The Federal Council's plan is to ratify the Council of Europe's Framework Convention on AI and implement it through targeted, sector-specific rules — in healthcare, transport, anti-discrimination — rather than a single omnibus statute, with the FDJP drafting implementing legislation for consultation by the end of 2026. That's the proportionate approach: address demonstrated harms, such as discriminatory credit or hiring algorithms or unsafe medical deployment, where they actually arise, instead of imposing a compliance regime on every model regardless of risk.
This is a coherent bet, not a contradiction: build open, auditable public infrastructure with public money, and regulate private deployment narrowly where rights are actually at stake. It's a sharper split of responsibilities than the EU AI Act's approach, which burdens general-purpose model developers with transparency obligations regardless of downstream use. If Switzerland wants that split to hold up, it has to stop calling CHF 20 million in Apertus funding "sovereignty" while committing roughly seven times as much to a foreign cloud vendor — and instead defend the model for what it plainly is: a transparent, auditable public option that expands genuine choice in a market currently dominated by a handful of opaque foreign labs. That's a good enough reason to keep funding it.