A Swiss Canton Is Running Government Work on a Publicly-Built AI
A Swiss Canton Is Running Government Work on a Publicly-Built AI — the US Export Ban Just Proved Why That Matters
In March 2026, the Canton of Ticino became the first government in the world to deploy a nationally-built open AI model for official work. Three months later, the US government switched off Anthropic’s models for every non-American overnight. Switzerland noticed.
How this was reported: the Ticino deployment facts in this article come from official press releases published on March 17 2026 by the Swiss National Supercomputing Centre (CSCS) and EPFL — both primary institutional sources, read in full. Quotes attributed to cantonal officials are reproduced verbatim from those documents. The Meditron/CHUV medical testing claim comes from a secondary Italian-language source and is flagged as such; Wangdoo has not independently verified it from a CHUV or Swiss AI Initiative primary publication. The Anthropic export ban facts are confirmed from Engadget’s reporting dated June 12 and June 30 2026. No interviews were conducted. This is not legal advice.
On March 17 2026, the Swiss National Supercomputing Centre published a press release that went almost completely unnoticed outside Switzerland. The Canton of Ticino’s Computer Systems Centre (CSI) had installed an AI translation tool running entirely on hardware inside the canton’s own data centre, using a Swiss-built open AI model called Apertus as its engine. Around 100 cantonal employees started using it for translating sensitive government documents between Italian, German, French, Spanish, Romanian and Ukrainian — with no data leaving Swiss infrastructure at any point. It was the first real government deployment of a nationally-built public AI model in production use anywhere in Europe.
Nobody outside Switzerland paid much attention. Then, on June 12 2026, the US Department of Commerce ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models for all foreign nationals. Unable to verify nationality across shared cloud infrastructure, Anthropic disabled both models worldwide. Every European organisation running on those models lost access without warning. Switzerland’s Canton of Ticino kept working.
The official Apertus launch video published September 2 2025 by EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) — the institutions that built the model. Not a Wangdoo production.
What Apertus Is
Apertus — Latin for “open” — is a large language model developed by the Swiss AI Initiative, a collaboration between EPFL, ETH Zurich, and CSCS. It was released in September 2025 and is available in multiple sizes from 0.5 billion to 70 billion parameters. It was trained on 15 trillion tokens across more than 1,000 languages, with 40 percent of its training data non-English by design — a deliberate choice for a country with four official languages and a government that routinely handles documents in several more.
Three things distinguish it from other open models. First: the training data, architecture, weights, and training recipes are all published. When Meta describes Llama as “open,” it means the model weights are available; the training data is entirely opaque. Apertus publishes the full pipeline, including reconstruction scripts. Second: it was built from the start to comply with European copyright and data protection law, respecting machine-readable opt-out requests from websites retroactively. Third: it is released under Apache 2.0 — no revenue caps, no “community licence” that disappears once your company grows past a certain size. Once you have the weights, nobody can revoke access.
It is not the most capable model available. On standard English-language benchmarks — complex reasoning, advanced coding — Apertus 70B trails GPT-5.5 and Claude Fable. That is not a surprise and the Swiss AI Initiative has not claimed otherwise. The model’s strengths are different: openness, regulatory compliance, multilingual breadth, and the ability to run entirely on your own infrastructure. Whether those properties matter depends on what you are trying to do.
Why Ticino, and Why This Deployment Is Significant
Ticino is one of only two Swiss cantons with Italian as an official language — the other being Grisons, where Italian sits alongside German and Romansh. As the only canton where Italian is the sole official language, Ticino handles a continuous volume of documents that need translation — between Italian, the languages of neighbouring cantons, and the languages of its diverse resident and migrant population. Many of those documents contain personal or legally sensitive information.
Until March 2026, the CSI used a translation tool built on Mixtral, a model from French company Mistral AI, running inside Artificialy’s “PrivateAI Suite.” The switch to Apertus was not primarily about capability. The canton’s head of systems and user support, Rudi Belotti, explained the reasoning directly in the EPFL press release: “As a public administration, we feel obligated to use ethical software applications. With Apertus we can be sure the model was trained in Switzerland and in accordance with the highest ethical standards, meaning it uses data that were not proprietary or copyright-protected but released for AI training.”
The technical setup reinforces the point. GPU processors were physically installed in CSI’s own data centre specifically to run local AI workloads. The model runs on-site. The translation requests never touch an external server. Luca Gambardella, CEO and co-founder of Artificialy and an AI professor at the University of Italian Switzerland, noted that despite using the smaller Apertus 8B model rather than the 70B version, fine-tuning for translation tasks produced strong results. According to Artificialy’s own benchmark tests — not an independently audited figure — the fine-tuned Apertus 8B reached 94 percent accuracy on translation, up from 90 percent before fine-tuning and above the 91.4 percent scored by Meta’s LLaMa Scout in the same comparison. “The good benchmark results were a prerequisite for our decision to use Apertus,” Gambardella said.
What the Export Ban Changed
The June 12 export control order was unprecedented. It was the first time the US government had used export controls to shut down a major AI lab’s deployed models in active production use. Anthropic, unable to verify the nationality of users across shared cloud infrastructure, disabled Fable 5 and Mythos 5 globally rather than selectively. European companies that had integrated those models into their products had no transition period and no warning. Access was there; then it was not. The ban was lifted on June 30 2026 — eighteen days later — after negotiations between Anthropic and the Commerce Department. That restoration does not undo the precedent. It confirms it: a deployed commercial AI model can be switched off by a government order with no warning, and restored only after the company satisfies government conditions. The eighteen days of downtime happened.
The reasoning behind Ticino’s Apertus deployment — that sensitive government work should not depend on infrastructure controlled by a foreign company, subject to a foreign government’s policy decisions — was not an abstract concern in March 2026. It became a concrete one in June. The canton had already solved this problem for its own use case before the export ban demonstrated the problem existed at scale.
What Apertus is and is not designed for
Apertus is not positioned as a frontier model, and the Swiss AI Initiative has not claimed otherwise. On standard English-language benchmarks — reasoning, coding — Apertus 70B performs at a level comparable to the Llama 3.1 70B generation, which trails current frontier models like GPT-5.5 and Claude Fable. It has no multimodal capability in its current release — text only, no image or audio input. Initial tests at the September 2025 launch found weaker-than-expected results for Swiss German and Romansh specifically, despite both being included in training data; the Swiss AI Initiative’s own technical paper identifies multilinguality improvements as an active development priority. The model’s design strengths lie elsewhere: full training transparency, EU AI Act compliance by design, Apache 2.0 licensing with no restrictions, and the ability to run entirely on your own infrastructure. Whether those properties outweigh the capability gap depends entirely on what you are building and what risks you are managing.
The UBS Counterpoint — and Why It Matters
Switzerland is not moving uniformly toward sovereign AI. UBS, the country’s largest bank, has not adopted Apertus. It is working with OpenAI and Microsoft. The Swiss Bankers Association has said Apertus has “great long-term potential” for banking compliance given its alignment with Swiss data protection and banking secrecy rules, but that potential has not yet translated into adoption by the major institutions who would benefit most from it.
This tension — a cantonal government running sovereign AI on public infrastructure while the country’s largest private bank continues paying American hyperscalers — is the more honest picture of where European sovereign AI actually stands in mid-2026. The export ban created urgency. It did not create universal adoption. The organisations most exposed to the risk are often the ones most deeply embedded in the infrastructure that creates it.
What Comes Next for Apertus
Since this article was first drafted, a primary source has emerged for part of the Meditron story. EPFL published a press release confirming MeditronFO — a framework that medicalises fully open LLMs including Apertus. The Apertus-70B-MeditronFO model improved performance on medical exams by 6.6 percentage points over the base Apertus model, per EPFL’s own reporting. The specific claim about clinical testing at Lausanne University Hospital (CHUV) from May 2026 remains from a secondary Italian-language source not yet confirmed by a CHUV or EPFL clinical primary publication; Wangdoo will update when that confirmation is available.
A new model version, Apertus v1.1 at 4 billion parameters, was released in June 2026 and is small enough to run on a consumer GPU. The full model family now spans five sizes from 0.5B to 70B, all available under Apache 2.0 via the swiss-ai organisation on Hugging Face. Enterprise deployment is available through Swisscom’s sovereign Swiss AI platform for organisations requiring guaranteed Swiss data residency.
The UAE is taking a different approach to the same question — building a federal AI authority with legislative and enforcement reach rather than investing in a sovereign model. Both bets are responses to the same underlying dynamic: the realisation that depending on a single provider’s goodwill, or a single government’s policy decisions, is infrastructure risk, not just procurement preference.
My Take — Mr Wangdoo
Reading through the CSCS and EPFL documentation rather than from any direct experience deploying Apertus: what strikes me about the Ticino case is how undramatic it was. A cantonal IT team made a pragmatic decision — our documents are sensitive, our data should stay here, the training data of our model should be something we can actually verify — and quietly installed it. No press fanfare. No sovereignty manifesto. Just a procurement decision grounded in compliance and ethics obligations that public administrations actually have.
The export ban then arrived and made that undramatic decision look prescient. The question for European organisations now is whether the export ban is treated as a one-time anomaly or as evidence that the underlying risk was always there. The Ticino deployment suggests some governments were already treating it as the latter, before the evidence arrived to confirm it.
Frequently Asked Questions
What is Apertus and who built it?
Apertus is an open-source large language model developed by the Swiss AI Initiative — a collaboration between EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS). It was released in September 2025, trained on 15 trillion tokens across more than 1,000 languages on the Alps supercomputer in Lugano, and released under Apache 2.0. The full training pipeline — architecture, weights, data, and recipes — is publicly available.
Why did the Canton of Ticino switch from Mixtral to Apertus?
Two reasons, both stated directly by the cantonal IT head Rudi Belotti in the official CSCS/EPFL press release. First, the canton’s obligation to use ethical software — Apertus was trained exclusively on publicly available, non-copyright-protected data, verified to Swiss standards. Second, sovereignty: running Apertus on hardware physically inside the canton’s own data centre means sensitive document data never leaves Swiss infrastructure, which cloud-based tools cannot guarantee regardless of their privacy policies.
Is Apertus as capable as GPT or Claude?
No, and the Swiss AI Initiative has not claimed otherwise. Apertus 70B performs at a level comparable to other open models in the 70B class, which means it does not match current frontier models like GPT-5.5 or Claude Fable on English reasoning, coding, or multimodal tasks. Its documented strengths are multilingual coverage, full training transparency, EU AI Act compliance by design, and Apache 2.0 licensing with no usage restrictions. The Ticino deployment chose it for those properties, not for raw benchmark performance.
What did the US export ban in June 2026 have to do with Apertus?
Nothing directly — Apertus is Swiss, not subject to US export controls, and runs on European infrastructure. The connection is the argument it validates. When the US ordered Anthropic to suspend Fable 5 and Mythos 5 for all foreign nationals on June 12, every European organisation using those models lost access without warning. The case for sovereign AI — build on infrastructure that cannot be switched off externally — shifted from policy theory to documented recent fact. Ticino had already acted on that argument three months earlier.
Can any organisation use Apertus?
Yes. Apache 2.0 licensing means unrestricted commercial use, modification, and redistribution. All model sizes are available on Hugging Face under the swiss-ai organisation. The 4B and 8B models are practical for fine-tuning on a single GPU. Enterprise deployment with Swiss data residency guarantees is available through Swisscom’s sovereign AI platform.
Sources
- Apertus powers in-house AI translation for Ticino — March 17 2026 — Swiss National Supercomputing Centre (CSCS) (primary source, read in full)
- Apertus powers in-house AI translation for Ticino — March 17 2026 — EPFL (primary source, read in full)
- Artificialy delivers secure AI translation for the Canton of Ticino — Artificialy (company that built and fine-tuned the deployment)
- EPFL launches the world’s first fully open medical LLMs — MeditronFO including Apertus-70B — EPFL (primary source, June 2026)
- US government allows Anthropic to redeploy its Mythos and Fable AI models — Engadget, June 30 2026 (export ban context)
