An open model with a government pedigree
On July 24, 2026, ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5, the second major version of Switzerland's fully open large language model. The update adds multimodal understanding — text, images and audio — plus stronger reasoning, better tool use, and a context window that quadruples to 262,144 tokens (CSCS). It was trained on Alps, the CSCS supercomputer in Lugano, using more than 20 million GPU hours a year allocated to the Swiss AI Initiative (ETH Zurich). The whole model suite ships under Apache 2.0 — weights, training data documentation and code all public.
That matters because Apertus is not a lab curiosity. It is public infrastructure, built with a specific political purpose: as ETH Zurich's Imanol Schlag put it, "Apertus is not about competing with frontier models from private companies, but about providing trustworthy, transparent AI" (CSCS). It is the clearest test case yet of a genuinely different national AI strategy: instead of writing a comprehensive rulebook first, Switzerland is trying to build the sovereign capacity to use AI on its own terms, and regulate narrowly around it.
The road not taken
The contrast with Brussels is deliberate. On February 12, 2025, the Federal Council decided against a horizontal, EU AI Act-style law. Instead it chose a sectoral approach — targeted amendments to existing laws in areas like healthcare and transport, plus ratification of the Council of Europe's Framework Convention on Artificial Intelligence, which Switzerland signed on March 27, 2025 (Federal Council decision, admin.ch). The stated objectives were threefold: strengthen Switzerland as an innovation location, safeguard fundamental rights, and build public trust in AI. A consultation draft implementing the Convention's transparency, non-discrimination and supervision requirements is due by the end of 2026.
The steelman for the EU's approach deserves a fair hearing. A single, comprehensive statute gives every company one compliance target instead of a patchwork of sectoral rules layered onto existing law, and it forces risk classification and fundamental-rights safeguards into every deployment before harm occurs, not after. For a bloc negotiating with hyperscalers as a unified market of 450 million people, that leverage is real. Switzerland, population 9 million, has no comparable leverage to extract concessions from Big Tech through regulation alone — which is precisely why its bet is different: build capability first, and let narrow, sector-specific rules follow the actual risks that show up in healthcare, transport and financial services, rather than pre-empting hypothetical ones across the whole economy.
The sovereignty the money can't yet buy
The honest complication is that Switzerland's compute bet is still underfunded relative to its own ambitions. The ETH Board committed CHF 20 million to the Swiss AI Initiative for 2025–2028, on top of roughly CHF 100 million already sunk into Alps (SWI swissinfo.ch). ETH's Alexander Ilic estimates the country needs 20 to 100 times more computing power within a decade to stay meaningfully sovereign, and EPFL's Marcel Salathé has said flatly that current investment falls well short of what sovereignty requires. Alps itself, like nearly every AI supercomputer on Earth, runs on Nvidia GPUs — a dependency no domestic funding round erases. The Swiss Science Council's May 12, 2026 recommendation for a national AI compute strategy makes the same point institutionally: project-based grants "cannot build sustainable capacity," and Switzerland needs a standing, tiered infrastructure commitment, not one more time-limited initiative (Swiss Science Council, admin.ch).
That gap is a legitimate criticism, but it argues for scaling up the compute bet, not abandoning it for a heavier regulatory one. A small country cannot regulate its way to frontier AI capacity; it can only build, share or rent it. Apertus's real innovation isn't the model weights — 8B and 70B parameter open models exist elsewhere — it's the demonstration that publicly funded, fully transparent training at national scale is operationally possible, auditable end to end, and already useful enough that a newsroom like Bajour is building on it because "we know what we're building on" (CSCS).
What should follow
The policy lesson generalizes beyond Switzerland. Countries worried about AI dependency on a handful of US and Chinese labs have two levers: regulate access, or build alternatives. Regulation without capability just changes who controls the choke point; it doesn't remove it. Switzerland's sectoral-plus-compute model is a more honest bet — proportionate rules where real harms are concrete (finance, healthcare), paired with sustained public investment in the infrastructure that makes genuine alternatives possible. The researchers proposing a "federation of sovereign AIs" — smaller nations pooling compute rather than each chasing hyperscaler-scale budgets alone — are describing the only version of this strategy that plausibly closes Ilic's 20-to-100x gap. Switzerland built the demo. Whether Bern is willing to fund the follow-through, rather than let Apertus 1.5 become a well-reviewed pilot that never scales, is now the actual policy question.