Giotto Launches Swiss Sovereign AI Models

Giotto, an AI lab in Lausanne, has launched two reasoning models that it says can give European users strong AI without relying on American cloud companies. Giotto Alto, its flagship, runs on eight Nvidia B300 GPUs, and the company says it is “made available for free to everyone”. Giotto Solo is designed to run on a single GPU, and Giotto calls it “the smartest model to run on one GPU”.

The launch was announced in a release on 27 October. Alongside the models, Giotto released apps for iOS and Android and a free anonymous chat that needs no account. The company says its cloud services are GDPR-aligned and run without US cloud infrastructure.

In brief

Giotto Alto runs on eight Nvidia B300 GPUs; Giotto Solo runs on one. Both are offered through European cloud or on a customer’s own hardware.
On Giotto’s own benchmark tables, Alto trails the top models it lists on two hard reasoning tests.
The release describes a flat-fee API; Giotto’s website lists per-token prices.

Who is behind Giotto

Giotto was founded in Lausanne in 2017 by mathematics graduates of EPFL, the Swiss federal technology institute. According to the release, it has worked on machine-learning interpretability, language models and reasoning systems. The company says its entries placed 11th in the 2024 ARC-AGI competition and second in 2025, among roughly 1,500 teams. ARC-AGI is a contest of visual puzzles designed to test whether AI systems can reason about problems they have not seen before.

The lab frames the launch as a mission rather than a product release. “Europe must offer more than a different flag on the same dependency,” said Aldo Podestà, Giotto’s chief executive and co-founder. “Sovereignty has to belong to the people and organisations using AI, and not just to the companies building it.”

He tied that to efficiency. “When intelligence requires infrastructure only a handful of companies can afford, control remains a privilege of scale,” Podestà said. “Better engineering can change who gets to hold that power.”

Three ways to run the models

Giotto offers both models in three ways. The first is a shared cloud, hosted and operated in Europe, which is free to use. The second is a dedicated cloud, rented monthly from a single GPU upwards, in isolated environments at partner datacentres in Switzerland or elsewhere in Europe; the price includes the software licence. The third is installation on a customer’s own compatible hardware through partner integrators, from a one-GPU workstation for Solo to multi-GPU server racks for Alto. Those on-premises deployments are licensed annually or over several years, priced by the number and type of GPUs.

For developers, the release says shared-cloud API access uses “a flat fee for all-you-can-use requests”, limited to one request at a time, with token packages for running several at once. Giotto presents that as a way of “avoiding uncapped per-token pricing”. Its API page, however, lists Alto token pricing of €2.10 per million input tokens and €10.50 per million output tokens. The two have not been reconciled publicly.

What Giotto’s benchmark tables show

The company’s website publishes scores for both models. All are Giotto’s own figures, and no testing method has been published alongside them.

On GPQA-Diamond, a set of graduate-level science questions, Giotto lists Alto at 91.9%, behind the 96.1% it gives OpenAI’s GPT 6 Astra and the 95.3% for Google’s Gemini 3.8 Flash. On Humanity’s Last Exam (HLE), a hard expert-level test, it lists Alto at 52.0%, against 59.1% for Anthropic’s Claude Fable 5.1 and 54.7% for GPT 6 Astra, but ahead of Gemini 3.8 Flash at 47.8% and Mistral Large 4 at 35.0%. Alto is shown scoring 100% on the AIME 2026 maths test.

For Solo, Giotto lists 88.3% on GPQA-Diamond and 37.5% on HLE, ahead of the 84.3% and 26.5% it gives Google’s Gemma 4 31B and the 80.1% and 19.0% for OpenAI’s GPT-OSS-120B.

A sovereignty pitch built on efficiency, with gaps to fill

Giotto’s argument is different from most European sovereign AI efforts. Governments have tended to fund national compute, as the UK did with its £500 million sovereign AI fund, or back partnerships between big labs and institutions, such as CERN’s agreement with Mistral. Giotto’s bet is that control comes from models small enough to move: if Solo really does run well on one GPU, a hospital or ministry could keep it on its own premises rather than sending data abroad.

That case rests on numbers only Giotto has produced so far. Its benchmark tables compare Alto with a different set of rival models on different rows, and give no settings or test dates, so they are not a like-for-like league table. Even on its own figures, Alto is a strong contender rather than a leader on the hardest reasoning tests. Independent leaderboards will matter more than the launch claims.

Several basics are also missing. Neither the release nor the website gives a parameter count, describes the training data or says what licence governs the models customers download to run themselves. That contrasts with labs such as Italy’s Domyn, which published open weights for its Edge model. “Sovereign” software that a customer can run but cannot inspect leaves some of the dependency Podestà says Europe must escape. The pricing mismatch between the release and the API page is a smaller point, but buyers will want it settled before signing up.