Swiss 'sovereign' Giotto model inherits Kimi weights

What launched?

Giotto.ai SA announced two models on October 27 (press release). Alto is the flagship and, according to the company, runs on eight Nvidia B300 GPUs. It is free on Giotto's shared European cloud and comes with iOS and Android apps. Solo is built for a single GPU and targets on-prem deployments. In its own announcement Giotto calls them "the smartest European sovereign AI models on the market", and the release stresses no dependency on US cloud providers.

No parameter count, context window or model card was published. The weights are not open and there is no Giotto model or config on Hugging Face, so the architecture cannot be checked from the outside.

How the Kimi base surfaced

An X user asked the Giotto chatbot in Chinese whether it was a Kimi model. According to the shared screenshot, it answered yes, developed by Moonshot AI and currently running as the "giotto-model" service (post). The post went up at 17:44 CET, about an hour and a half after the launch.

Eight minutes later co-founder and CEO Aldo Podestà replied: "we do inherits some weights from kimi as well", adding that the overall model was retrained as part of a larger effort, plus proprietary test-time compute (TTC) (reply). He later wrote that there is "much more than kimi involved". He did not say which Kimi version was used or which weights were inherited. It is also unclear whether this applies to Alto only or to Solo as well.

The approach itself was not secret. In July the NZZ reported that Giotto post-trains open models, mostly Chinese ones, instead of pretraining its own (NZZ). Kimi was not named then, and the launch materials do not mention it either.

What the benchmarks show

Giotto's own table lists Alto at 100.0% on AIME 2026, 91.9% on GPQA-Diamond and 52.0% on HLE (giotto.ai). The same table puts Kimi K3 (max) at 46.9% on HLE, so the company ranks its model above the family it partly builds on. The Math-500 comparison uses GPT 5 (high) and Claude Sonnet 4, which critics called outdated reference points.

Source: Giotto. The company's benchmark table for Alto and Solo. All numbers are self-reported.

There is no independent evaluation, and the license terms make one harder. As of Wednesday morning Giotto's EULA still barred users from benchmarking the model for publication without prior written consent, and from reverse engineering its parameters or architecture (EULA). Responding to criticism, Podestà called it a fair point, said they would amend it and that they are not against independent benchmarking (post).

Why it matters

Building on open weights is standard practice, and Kimi's license explicitly allows fine-tuning and derivative models. The Kimi K3 license only requires displaying "Kimi K3" above 100 million monthly active users or 20 million dollars in monthly revenue (Hugging Face). The issue is not legal, it is about disclosure.

The irony: on October 19 Podestà published "Open Source Is a Trojan Horse", arguing that open models are sold to Europe as sovereignty while often meaning dependence. Giotto's own definition of sovereign is deployment on the infrastructure of your choice. That may hold for data handling, but not for where the model comes from.

The practical takeaway: when buying a "sovereign" model, ask what the base model is, which license it carries and whether you may test it independently. A vague answer is a signal in itself.