By Stuart Kerr, Technology Correspondent, LiveAIWire
The Apertus AI model has passed one million downloads since its September 2025 launch, and ten months later it is no longer a proof of concept sitting on a shelf. Researchers at EPFL, ETH Zurich and the Swiss National Supercomputing Centre built Apertus as a fully open large language model, with published weights, architecture and training data, at a moment when every leading commercial model remains a closed system guarded as a trade secret. What has changed since launch is not the model’s headline ambition. It is the evidence, good and bad, of what happens when a government-funded open model actually gets deployed against real users rather than benchmark leaderboards.
The Apertus AI model comes in two sizes, eight billion and seventy billion parameters, trained on 15 trillion tokens across more than 1,800 languages using the Alps supercomputer in Lugano, powered by Swiss hydroelectricity. It is hosted on Hugging Face, through Swisscom’s sovereign AI platform, and increasingly through Infomaniak’s Swiss cloud infrastructure. The Swiss government has committed 20 million Swiss francs to the Swiss AI Initiative through 2028, funding that treats Apertus as long-term public infrastructure rather than a one-off research release.
What Real-World Deployment of the Apertus AI Model Actually Shows
The clearest independent evidence of how the Apertus AI model performs in production comes from Liip, a Swiss digital agency that has run Apertus alongside GPT-4o-mini on ZuriCityGPT, a public-facing chatbot answering questions about the City of Zurich, since the model’s release. Over eight months and more than 42,000 real conversations, Liip found that both models produce an acceptable answer at a similar rate, 82 percent for Apertus against 81 percent for GPT-4o-mini. The gap opens when the bar rises to genuinely good answers, where GPT-4o-mini scores 72 percent against 55 percent for Apertus, and on faithfulness to source material, where GPT-4o-mini averages 0.79 against 0.50 for Apertus.
Liip’s data also complicates any simple story about which model is better. The Apertus AI model outperformed GPT-4o-mini on at least one real question, providing a detailed, correctly sourced answer about Zurich’s waste disposal rules where GPT-4o-mini declined to answer at all. The Apertus AI model also struggled with basic instruction-following, at one point failing to answer a question posed in French despite being asked to do so directly, and returned a stale 2013 population figure for Zurich where GPT-4o-mini gave the current number.
Liip’s own verdict is not that Apertus has caught up to commercial models. It is that the quality gap is narrowing month over month, from a Match Rate of 59 percent in October 2025 to 63 percent by June 2026, without a major model update in between.
Language switching remains one of the model’s clearest weak points, and Liip’s testing found Apertus declined to answer a direct question posed in French, saying it could not answer questions about its own linguistic capabilities, where GPT-4o-mini responded fluently without hesitation. Liip has since observed improvement in more recent testing, and the upcoming Apertus 1.5 release is expected to bring gains in multimodal support, agentic tool calling, and specifically better training for regional Swiss languages, an area where the current version’s instruction-following score of 44 percent on the IFEval benchmark trails comparable models scoring around 58 percent.
From Chatbot Demo to Hospital Emergency Room
The most consequential test of the Apertus AI model in 2026 is not a city chatbot but a hospital. Lausanne University Hospital, known as CHUV, began testing Meditron, a specialised medical model built on the Apertus base, in its emergency room starting in May 2026. More than 300 health professionals evaluated Meditron using hypothetical clinical cases in 2025 before the live trial began, and researchers involved in the project found that more than 70 percent of doctors across the countries participating in the Meditron project already use commercial tools such as ChatGPT in clinical practice today, whether or not their hospital has approved that use.
That existing behaviour is precisely the risk Meditron is designed to address. Mary-Anne Hartley, the EPFL professor coordinating the Meditron project, has argued that many commercial models used informally in medicine are unreliable for clinical work because privacy requirements prevent them from being trained on real, sensitive patient data. An open model that hospitals can run locally, without sending patient records to an external server, can be tailored to a hospital’s specific clinical context in ways a closed commercial API cannot, at least not without the hospital surrendering exactly the data control that made open deployment attractive in the first place.
The Meditron project matters beyond its own clinical results because it demonstrates a pattern the Apertus AI model was always designed to enable: specialised, vertical models built on top of a shared open foundation, tailored to a specific institution’s data and regulatory constraints, rather than a single general-purpose product trying to serve every use case at once. That is a meaningfully different strategy from how commercial frontier labs currently compete, and its success or failure in a live hospital setting will say more about the practical value of open-weight AI than any further benchmark comparison could.
What Made This Level of Transparency Possible
The institutions behind the Apertus AI model bring together research capacity, public funding and computational infrastructure in a combination difficult to replicate outside a well-funded academic and public sector context. ETH Zurich and EPFL rank consistently among Europe’s leading technical universities, and the Swiss National Supercomputing Centre provides compute infrastructure to train models at a scale that would otherwise require the resources of a large technology company. The Swiss federal government’s investment in digital sovereignty created the conditions for a project like this to proceed without commercial pressure to keep results proprietary, a structural advantage that most countries attempting to build a sovereign AI model do not have.
Why Switzerland Chose a Different Regulatory Path Than the EU
Switzerland’s relationship to the Apertus AI model cannot be separated from a deliberate regulatory choice made in the same period. Rather than adopt an EU AI Act style law imposing binding requirements on high-risk systems, Switzerland decided to ratify the Council of Europe’s AI Convention and adapt national law around it, with the government set to propose binding and non-binding measures in sensitive sectors such as healthcare and autonomous transport by the end of 2026.
LiveAIWire’s coverage of how the AI regulatory landscape split between the EU and the US in 2026 found a global pattern of governments choosing sharply different enforcement models for broadly similar concerns, and Switzerland’s lighter-touch approach, paired with direct public investment in an open model, is a third distinct path alongside the EU’s binding statute and Washington’s attempt to suppress state-level rules entirely.
That choice has not been free of contradiction inside Switzerland’s own AI strategy. LiveAIWire’s earlier reporting on Meta’s rejection of the EU’s voluntary AI code found that major commercial developers are already split on how far they will go to accommodate government oversight, even under a voluntary framework far less demanding than a binding statute. Switzerland’s own government is spending 140 million Swiss francs over the next two years renewing its Microsoft licences, even as officials describe reducing dependence on foreign technology suppliers as a strategic priority, an inconsistency Swiss AI researchers themselves have pointed out directly.
The Sovereignty Argument Extends Well Beyond One Model
The logic behind the Apertus AI model, that relying on AI systems built and controlled elsewhere creates a strategic vulnerability, is the same logic playing out at much larger scale elsewhere in the AI industry. LiveAIWire’s reporting on the AI export controls that shut down Anthropic’s most advanced models worldwide in June 2026 found that a single Commerce Department directive could remove a piece of critical AI infrastructure from allied nations overnight, a demonstration of exactly the dependency risk that motivated Switzerland to fund a fully open, domestically controlled alternative in the first place.
LiveAIWire’s coverage of the AI Cold War between the United States and China found the same sovereignty anxiety shaping chip export policy at a geopolitical scale far beyond what any single open-weight model can address on its own. Switzerland’s own answer to that anxiety extends past Apertus itself. The government has also proposed hosting a global AI Action Summit in Geneva in 2027, explicitly framing the country’s neutrality and diplomatic tradition as an asset for mediating between competing national approaches to AI governance, the same way Geneva has long hosted humanitarian and disarmament negotiations.
What This Means for Anyone Evaluating Open Versus Closed AI
The practical lesson from ten months of Apertus deployment data is not that open models have caught up to commercial frontier systems, because on Liip’s own numbers they clearly have not on every metric that matters. It is that the gap has become genuinely a trade-off rather than a foregone conclusion. Liip’s own advice to other organisations considering a similar move is direct: start with the specific use case, not a general preference for openness or performance, because a well-curated internal knowledge base narrows the quality gap in ways a broad public chatbot never will.
For governments and institutions specifically weighing data sovereignty against raw capability, as CHUV is doing in its emergency room and the Canton of Ticino has already done for official translation, the Apertus AI model offers something no commercial API can: the ability to run the entire system on infrastructure the institution actually controls, audited down to the training data itself. Whether that sovereignty is worth the measurable quality gap Liip’s data documents is not a question with one right answer. It depends entirely on what an organisation is actually trying to protect, and how much it is willing to pay, in performance terms, to protect it.
The Canton of Ticino’s own deployment is instructive precisely because it is unglamorous. Rather than a flagship national showcase, a regional government simply needed a reliable translation tool for official documents and found that a fine-tuned version of the smaller Apertus-8B model met that specific, bounded need without requiring any data to leave Swiss infrastructure. That kind of quiet, narrow deployment, solving one well-defined problem rather than attempting to match a frontier chatbot across every task, is likely to be where the Apertus AI model accumulates its most durable value over the next several years, even as the headline comparisons with commercial models continue to generate the bulk of public attention.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.