TLDR: With its open-source hosting and new consumer product, Euria, Infomaniak is positioning itself as Switzerland’s answer to ChatGPT and similar services—transparent, GDPR-compliant, and utilizing waste heat in Geneva. The Swiss language model Apertus, developed by ETH Zurich, EPFL, and CSCS, is the flagship of digital sovereignty. However, the same model toolkit also includes Chinese open-source models such as Qwen and Kimi-K2.6, and the NZZ am Sonntag revealed that Euria adopts the Chinese government’s stance on Taiwan. At Infomaniak, sovereignty is not an illusion, but it is a choice that must be actively made—not a default state.
Two products, one company, one promise
Infomaniak, the Geneva-based cloud provider, has been pursuing two parallel AI initiatives since the start of this year. The first is open-source hosting at infomaniak.com/hosting/ai-services: a REST API that allows developers and businesses to access eight large language models, several reranking and embedding models, Whisper for transcription, and Photomaker and Flux for image processing. A free allowance of one million credits allows users to get started with no obligation; after that, billing is based on usage.
The second track is Euria, which has been live since December 2025: an end-user chatbot that is free in its basic version and offers paid upgrades via My-kSuite (19 CHF/year) and kSuite Pro (22.90 CHF per user/year). Euria transcribes, summarizes, analyzes images, independently performs web searches, and offers a «discreet mode» with no data storage. Its target sectors are specifically healthcare, education, law, finance, and public administration—industries where data protection isn’t just a nice-to-have but a legal requirement.
Both products make the same promise: data centers in Switzerland, processing in accordance with Swiss and European law, and no training using user data. This is credible—as long as you don’t ask which model is actually responding in the background.
Apertus: The Model at the Heart of the Sovereignty Debate
Apertus is the real story behind the story. Developed by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre, with Swisscom as a strategic partner and funded by more than 10 million GPU-hours, the model was released on September 2, 2025. It comes in two sizes—8 and 70 billion parameters—and was trained on 15 trillion tokens in over a thousand languages, including Swiss German and Romansh. Forty percent of the training data is non-English.
What matters is not the size, but transparency. The architecture, weights, training data, checkpoints, and methodology are fully open under the Apache 2.0 license, including compliance with European copyright law regarding data selection. Apertus is thus considered the first large model to meet the requirements of the EU AI Act from the ground up, rather than having been retroactively adapted. ETH Zurich itself compares the project to public infrastructure—roads, water, and electricity—only in the context of language models.
This transparency is not a minor issue. It is the difference between a model whose behavior can be understood and one that can only be evaluated based on its outputs.
The Model Mix: Where Sovereignty (Almost) Ends
This is exactly where things get a little tricky. In addition to Apertus, Infomaniak’s model lineup includes Kimi-K2.6 from Moonshot AI, Qwen 3.5 in two sizes from Alibaba, and other models developed outside of Europe. These models are powerful and cost-effective to run, which is why they end up in almost every open model toolkit, including those from competing providers.
The problem isn’t their origin per se, but rather what comes with it. The Center for European Policy Analysis has documented that Chinese open-source models such as Qwen and DeepSeek contain built-in behavioral guidelines, such as answering questions about China „positively and constructively, without criticism.“ When it comes to war-related topics such as the situation in Ukraine, some of these models unsolicitedly promote Chinese positions. Over 2,800 derivatives of Qwen are now circulating worldwide, often without developers realizing what biases they are inadvertently adopting.
This is precisely the pattern the NZZ am Sonntag observed with Euria: In a test, the chatbot referred to Taiwan as an „inalienable part of China“—that is, Beijing’s official position—rather than offering a neutral assessment. This finding is significant not because it shows that Infomaniak is lying, but because a product marketed as independent adopts the language of a foreign state on a geopolitically sensitive issue without users realizing it.
What this means for businesses and users
The practical implication is simple, but it’s rarely followed: With Infomaniak, sovereignty isn’t a default setting—it’s an option in the dropdown menu. Anyone working via the API can specifically request Apertus and thus remain within the European framework—from the training dataset to the response logic. Anyone using Euria’s standard chat leaves the model selection up to the platform, which may end up using a Chinese model depending on the query.
For companies in regulated industries, this means: Before using a „Swiss“ or „European“ AI tool, it’s worth asking the provider specifically which model is actually running for which type of query—and whether that can be specified. Digital sovereignty isn’t a label you buy—it’s a configuration decision you have to make. Infomaniak provides the infrastructure for this with APERTUS. Ultimately, it’s up to each user to decide whether to use it.
And that’s why I remain convinced: The future no longer belongs to typing and clicking, but to working together with intelligent systems. In this context, autonomy doesn’t come from a label of origin, but from knowing which model truly thinks for you—and from the conscious decision to choose that one. So if you want to talk, and if you want to work with me, feel free to reach out.
Disclaimer:
This article was compiled manually based on my own knowledge and supplemented by AI-powered research (Perplexity.Ai and Gemini.Google.com), and then simplified using Deepl.com/write. The text is then reviewed and critically evaluated by two people of my choice. The image is from AI-generated imagery (Ideogram/Adobe Firefly). This article is purely educational and does not claim to be exhaustive. Please let me know if you notice any inaccuracies; thank you.
Sources and Additional Information
Center for European Policy Analysis. (2026). Chinese AI models spread propaganda worldwide. https://cepa.org/article/chinese-ai-models-spread-propaganda-globally/
ETH Zurich. (September 2, 2025). Apertus: A fully open, transparent, multilingual language model [Press Release]. https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html
EPFL. (2025). Apertus: An open, transparent, and multilingual language model. https://actu.epfl.ch/news/apertus-ein-offenes-transparentes-und-mehrsprachig/
Infomaniak. (n.d.). AI service – open-source AI on demand via API, hosted in Europe. https://www.infomaniak.com/de/hosting/ai-services/open-source-models
Infomaniak. (n.d.). Euria – the sovereign AI. https://euria.infomaniak.com/
IT Magazine. (December 2025). Infomaniak Launches Euria, an AI Designed with Privacy in Mind. https://www.itmagazine.ch/artikel/86168/Infomaniak_bringt_auf_Vertraulichkeit_ausgelegte_KI_Euria.html
Big Data Insider. (n.d.). AI Assistant Euria Is Designed to Heat Homes. https://www.bigdata-insider.de/ki-assistent-euria-soll-wohnungen-beheizen-a-0bf9b69148ee9a4e779537d777a44dba/
NZZ am Sonntag. (February 2026). The Swiss chatbot Euria promotes the Chinese government's ideology. https://www.nzz.ch/nzz-am-sonntag/schweiz/die-schweizer-alternative-zu-chat-gpt-sagt-taiwan-ist-kein-unabhaengiges-land-es-ist-ein-unveraeusserlicher-teil-chinas-ld.1926690 [Original article behind a paywall; date reconstructed from a secondary source and not independently verified]
swissAI. (2026). Swiss AI Chats: Sixteen Options, Three Approaches. https://www.swissai.ch/blog/schweizer-ki-chats
Wikipedia. (n.d.). Apertus (LLM). https://en.wikipedia.org/wiki/Apertus_(LLM) [supplementary source, not a primary source]
