One year after its release, Switzerland's publicly funded Apertus language model has more adoption than its critics predicted. It also trails the best commercial systems, as its own supporters concede. The sensible reading is that the trade-off was chosen deliberately, and that Bern should keep backing the project without measuring it against frontier benchmarks.
What the one-year record shows
On 2 September 2026, Swissinfo's one-year review reported that Apertus has passed 4 million downloads on Hugging Face. Community lead Oleg Lavrovsky told the outlet he knows of at least 70 external deployments and more than 100 derivative versions.
Some of those derivatives are consequential. Singapore's SEA-LION project built on Apertus, and its February 2026 announcement of an 8-billion-parameter Apertus-SEA-LION model cited the open dataset and training pipeline as the reason. EPFL's MeditronFO clinical framework is also built on Apertus. In Ticino, the migration office uses the model to translate administrative documents.
On 24 July 2026, ETH Zurich, EPFL and the Swiss National Supercomputing Centre released Apertus 1.5. It was trained on the Alps supercomputer in Lugano and released under Apache 2.0. CSCS says the model now understands images and audio alongside text, and it reports improved reasoning, instruction following and tool use. It also introduced Apertus Mini, a suite of 16 compact models. Swissinfo reports that Apertus 2.0 is planned for 2027.
The strongest case against
Skeptics have a serious argument. Swissinfo quotes experts who say the 70-billion-parameter model remains well behind comparable open-weight competitors in programming, mathematics and reasoning. Federico Magnolfi of Artificialy estimates that Apertus 1.0 launched around two years behind the best US models. Public money spent on a model that trails the frontier looks hard to defend, especially when commercial and open-weight alternatives are free or cheap. Allocation of compute inside the Swiss AI Initiative is still debated, and every GPU-hour given to one project is unavailable to another.
That case is right about the gap and wrong about what the gap means.
Why the gap is not the test
Apertus was not built to win coding leaderboards. Swissinfo notes that transparency requirements limit the training data available to it compared with proprietary developers. That constraint is the product. A model whose data, code and training recipe are public can be audited, reproduced and adapted by a public administration, a hospital or a university without a licensing relationship or a foreign provider's terms of service. AI Singapore's choice of Apertus as a foundation is evidence that this has real value beyond Switzerland.
The deployments fit this. Translating administrative documents across four national languages, or adapting a model for regional languages, does not need the strongest reasoning model. It needs one that is good enough, inspectable and free to modify. The relevant question is whether open publicly funded infrastructure gives Swiss and international users options they would otherwise lack. On the evidence so far, it does.
The Swiss approach also avoids a common mistake in industrial policy. Apertus is a research-institution project on shared national compute, released under a permissive licence. It is not a state champion protected by procurement rules or by regulation. Anyone can fork it, compete with it or ignore it.
Regulation should stay light around it
That matters for the surrounding legal framework. On 12 February 2025, the Federal Council decided to ratify the Council of Europe's AI Convention and amend Swiss law accordingly. It said work would continue on sector-specific rules, for example in healthcare and transport, rather than a single EU-style AI Act. Sector-specific rules are the better fit for an open-model ecosystem. A horizontal regime with heavy obligations on general-purpose model developers would fall hardest on small academic teams releasing open weights, who are least able to absorb compliance costs.
Switzerland's regulators can be proportionate because the biggest transparency and rights risks in public-sector AI use are addressed through the state's own procurement and deployment choices. An open model that an administration can audit makes those choices easier to scrutinise.
What Bern and the Swiss AI Initiative should do
Three practical steps follow from the record.
- Publish compute allocation criteria. The debate over how Alps capacity is divided within the Swiss AI Initiative will not go away. Transparent, competitive allocation rules would make the case for continued funding and answer critics.
- Measure what Apertus is for. Report adoption, auditable deployments, derivative models and cost per public-sector task alongside benchmarks. A two-year lag on reasoning tests says little about whether a translation workflow in Ticino works.
- Keep regulation sector-specific and light on open weights. Consultation on implementing the Council of Europe convention should make sure any new duties do not push academic open-model developers out of the country.
Apertus will probably stay behind the frontier, because the frontier is funded at a scale no Swiss public programme can match. But the 1.5 release and the derivative ecosystem show that an open, publicly funded model can be useful. Switzerland should keep funding it and keep the rules around it proportionate.