Thomas Nedjar
All apps

Giotto’s sovereign AI from Switzerland: what is disclosed, and what is not yet

Published on 29 October 2026

On 27 October, Lausanne-based Giotto.ai launched two language models, Giotto Alto and Giotto Solo, and called them “the smartest European sovereign AI models on the market”. If you run a European business looking for an alternative to American and Chinese providers, the announcement deserves a careful read. The word sovereign carries a lot of weight in it. Here is what has been published, what has not, and what you can check yourself before trusting any model with your data.

What Giotto announced

Giotto.ai SA is based at Place de la Gare in Lausanne. According to its press release, it was founded in 2017 by EPFL mathematics graduates. It reports an 11th place in the ARC-AGI competition in 2024 and a 2nd place in 2025, out of roughly 1,500 teams.

Solo runs on a single GPU. Alto, the flagship, targets advanced reasoning and enterprise workloads and runs on eight Nvidia B300 GPUs.

There are three deployment options:

  • a shared cloud, hosted and operated in Europe, free for basic use;
  • a dedicated cloud in partner data centres in Switzerland or elsewhere in Europe, rented monthly;
  • on-premises installation through integrators, from a one-GPU workstation for Solo to multi-GPU racks for Alto.

The press release describes a flat-fee API with unlimited requests and one concurrent request. The API page also lists usage-based pricing for Alto: EUR 2.10 per million input tokens and EUR 10.50 per million output tokens. CEO Aldo Podestà sums up the ambition: Europe must offer “more than a different flag on the same dependency”.

The benchmarks, as published by the vendor

On its website, Giotto reports 91.9% for Alto on GPQA-Diamond, against 96.1% for GPT 6 Astra and 93.7% for Claude Fable 5.1. On Humanity’s Last Exam, Alto scores 52.0%, against 54.7% and 59.1%. On AIME 2026, it reports 100%. The same table puts Mistral Large 4 at 35.0% on Humanity’s Last Exam.

If confirmed, these numbers place Alto just behind the best closed American models and well ahead of Europe’s best-known lab on that test. They are vendor-reported. As of 29 October, I found no published independent evaluation.

What has not been published

The parameter count, the architecture, any base model, and the composition of the training data. The press release and the website say nothing on these points.

The weights are not published either. The EULA defines the model as “proprietary”, delivered as a container, and prohibits any attempt to derive its “source code, parameters or architecture”. On-premises deployment exists, under licence. The same text states that the model may include open source components governed by their own licences, listed in official documentation I have not reviewed.

That is the core of the sovereignty question. Hosting determines where your data is processed. Provenance determines what you are actually running: its biases, its gaps, and the legal terms of any underlying base. A model hosted in Lausanne still carries its origin.

Switzerland has already shown the other path. On 2 September 2025, EPFL, ETH Zurich and the Swiss National Supercomputing Centre released Apertus in 8 and 70 billion parameter versions, with weights, training data, code and documentation all public, under a licence that allows commercial use. Its training data covers more than 1,000 languages. Apertus trails frontier models on performance and is fully auditable.

What the contract says

The EULA published on Giotto’s site, October 2026 version, contains concrete commitments. Swiss law applies, with jurisdiction at the company’s registered office. The cloud service runs in data centres in Switzerland or the EU. Giotto commits not to use your inputs to train the model. Feedback you choose to send may be used to improve it.

The same EULA limits the licence to “non-commercial testing and evaluation purposes” and prohibits reverse engineering. The paid offers, dedicated cloud and on-premises, are presumably covered by separate contracts that I could not review. In practice, the free tier is for testing. Production use requires a contractual discussion.

Five questions for any “sovereign” provider

I apply the same grid to Giotto, to Mistral, and to an American model hosted in a European cloud region.

  1. Provenance. Trained from scratch or built on a base model? Which one, under which licence?
  2. Usage rights. Does the licence cover your real, commercial use?
  3. Data processing. Where, by whom, under which law? Is your data used for training? Is a GDPR-compliant data processing agreement available?
  4. Reversibility. Can you switch models without rebuilding your tools?
  5. Performance on your tasks. Twenty real cases from your business tell you more than a standardised exam.

On paper, Giotto answers the third question convincingly. The first remains open. The second depends on the enterprise contract. The last two can be tested in a few days.

How I would use it

I would open the free cloud with a set of tasks representative of your business, with no confidential data, and compare the output with the model you use today. Then I would ask the vendor, in writing, about the model’s provenance and the contract that applies to commercial use.

On the technical side, I would keep the rule I apply to all my business tools: an abstraction layer between your applications and the model, so you can switch providers without rebuilding. That is the optionality I wrote about with Kimi K3. One more credible European provider widens your options. How much that option is worth depends on what Giotto discloses about the origin of its models in the coming weeks.

Read also: Kimi K3: the US or China, the only question Europe still asks itself · AI agents in the enterprise

Thomas Nedjar
Thomas Nedjar
SEO/GEO & automation expert

Thomas Nedjar has spent fourteen years in digital acquisition and e-commerce. An entrepreneur, he founded and ran several e-commerce companies between France and Switzerland, with solid experience in international and intra-EU trade. He is now Senior SEO/GEO, AI & Tech Expert at Suisseo, and builds his own SEO tools, including SEO Cartograph (a crawler paired with an AI-driven SEO dashboard). A speaker (CVCI) and trainer, he shares his hands-on insights on SEO, GEO, automation and e-commerce here.

All articles · LinkedIn

Latest articles

You might also like

France
Suisse