Giotto, the "sovereign" Swiss AI, bans publishing tests without its consent

“Europe must offer more than another flag on the same dependency.” The line comes from Aldo Podestà, co-founder and CEO of Giotto, as quoted by ICTjournal. On October 27, his Lausanne-based lab unveiled two reasoning AI models, Giotto Alto and Giotto Solo. It presents Solo as the smartest model able to run on a single graphics card.

Founded in 2017 by former mathematics students of the École polytechnique fédérale de Lausanne (EPFL), Giotto promises AI hosted in Europe, GDPR-compliant and independent of American clouds. But researchers and internet users have asked a simpler question: how can anyone verify what it promises?

Two AIs: one free, one you can install yourself

Alto is the flagship model, designed for complex reasoning and professional use. It runs on eight Nvidia B300 graphics cards. It can be used for free on the web and in iOS and Android apps, even without creating an account. Solo, which is smaller, fits on a single graphics card: it targets organizations that want to run AI on their own machines.

Giotto offers three plans: a shared cloud operated in Europe, a dedicated cloud hosted by partners in Switzerland or the European Union, and an on-premises installation under an annual or multi-year license. API access is sold as a flat rate, but no price is published. According to ICTjournal, which draws on an interview given to Bilan, the lab is betting on compute at response time rather than on ever-larger models. Giotto discloses neither the number of parameters nor the architecture of its models, and does not publish their weights.

Performance that only Giotto has measured

The only scores available come from Giotto. Its press release highlights the team's rankings in the ARC-AGI competition (2^nd out of about 1,500 teams in 2025), not independent measurements of Alto or Solo. Critics noted that its charts compare math performance with Claude Sonnet 4, an Anthropic model released in May 2025.

Verification will be difficult. Researcher Alexander Doria, of the French start-up Pleias, pointed out on X clauses in the terms of use that prohibit reverse engineering the model and subjecting it to tests intended for publication without Giotto's written consent.

A Swiss AI, or a Chinese one under another name?

Another question: according to the news account AGTP, users got Alto to introduce itself as a Kimi model, built by China's Moonshot AI. An AI that gets its own identity wrong proves nothing, as these models often describe themselves inaccurately. But building a product on an existing open model is common and legitimate, as long as it is disclosed. Giotto has not responded publicly so far. Moonshot does publish high-performing open models, such as Kimi K3.1 Pro, close to the best American models according to an independent test.

What a user can do with it right now

For individuals, Alto can be tested without an account and at no cost. For a company, the sovereignty criterion is about hosting and the contract, not just the flag: ask which base model Solo is built on, who can audit its performance, and what the contract allows you to test. By comparison, Mistral took the opposite route by announcing open weights for Mistral Large 4.

What happens next is up to Giotto: publish the origin of its models, or let independent evaluators measure them.