Swiss Startup Giotto's Fine-Tuned Chinese Kimi Was Sold as Europe's Sovereign AI
Swiss startup Giotto released Alto and Solo as Europe's 'sovereign' AI, with data processed only in EU data centers. But during testing, Alto identified itself as Kimi, the Chinese model from Moonshot AI, and CEO Aldo Podesta acknowledged that the system runs on Kimi's borrowed weights, a ready-made model with additional fine-tuning. When evaluating a 'European' model, ask which base model is underneath.
AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
Swiss startup Giotto presented Alto and Solo as "Europe's sovereign AI": a development that does not depend on American big tech, with data and computation kept only in European Union data centers. A couple of days later, it emerged that Alto runs on Kimi, a Chinese language model from Moonshot AI. Put simply, it's like buying a finished car, swapping the badge, and selling it as your own design. Technically, this is called "borrowed weights." Weights are a trained neural network, and Giotto took someone else's instead of training its own from scratch.
What happened at launch
Giotto announced Alto and Solo as the smartest European development. The main argument was the rejection of American tech giants' infrastructure: data and computation, the company promised, stay exclusively within the European Union.
Then users started testing Alto. Asked what model it was, Alto replied that it was China's Kimi and even mentioned the "dark side of the moon." This is a reference to the developer, Moonshot AI. This happens when a model is built on someone else's weights and retains a "memory" of its original identity.
After the discussion on X, Giotto CEO Aldo Podesta acknowledged that Alto is indeed built on Kimi's borrowed weights. In his defense, he pointed to additional training and a "proprietary Turing Test Calculation technology." He did not explain what that technology is. Giotto also did not disclose which version of Kimi was used.
Why the story became especially awkward
Eight days before the incident, Podesta published an article titled "Open Source Is a Trojan Horse." In it, he sharply criticized European businesses for living off other people's open models and passing this off as sovereign technology. In effect, he had described his own scenario before he was caught.
Choosing a ready-made model as a foundation is not prohibited in itself and is common practice in the industry. Fine-tuning existing weights is cheaper than training from scratch, and many companies do it. The problem lies in the presentation: if a product is positioned as an independent European development but runs on a Chinese model underneath, the customer does not get what they were promised.
This is not an isolated case
A similar story already happened in September. Spain's Multiverse Computing introduced Quasar 438B as a leading European model. Only the day after the announcement did a caveat appear in a technical blog post: it is an optimized version of GLM-5.2, a Chinese model.
Two cases in a row show a pattern. A "European model" on the market can mean three different things: trained from scratch in Europe, fine-tuned in Europe on someone else's weights, or simply hosted in a European data center. For many tasks the differences are not critical, but marketing often blurs them.
What developers and product managers should do
If you are choosing a "sovereign" model for a bank, healthcare provider, or government client, do not stop at the slogan on the website. A few simple checks:
- Ask the vendor which base model sits underneath and which version it is.
- Clarify where computation physically takes place and where data is stored, and write it into the contract.
- Review the license of the original model: it may set conditions on commercial use and attribution.
- Ask the model several questions about itself. As the Alto case showed, such answers sometimes reveal its origins.
There is also an honest limitation. Based on available data, it is impossible to assess how much Alto differs from Kimi after fine-tuning, or whether that fine-tuning produced a real quality gain. Giotto has not disclosed these details. Until it does, the claim of "the smartest European model" should be treated as advertising, not as a verified fact.
- Jul 31, 2026Moonshot's Kimi K3: Why China Is Giving Away Its Best AI Models for Free
- Oct 10, 2026Alibaba Chairman Joe Tsai Calls Open Source Europe's Only Path to Tech Independence
- Nov 1, 2026Kimi scores 96.2% on a 1C database analytics test and emerges as a fallback if Claude access is blocked
- Apr 22, 2026Cursor admits its new coding model is built on China's Kimi from Moonshot AI
Key Questions Answered
What happened at launch?
Giotto announced Alto and Solo as the smartest European development. The main argument was the rejection of American tech giants' infrastructure: data and computation, the company promised, stay exclusively within the European Union. Then users started testing Alto. Asked what model it was, Alto replied that it was China's Kimi and even mentioned the "dark side of the moon." This is a reference to the developer, Moonshot AI.
Why the story became especially awkward?
Eight days before the incident, Podesta published an article titled "Open Source Is a Trojan Horse." In it, he sharply criticized European businesses for living off other people's open models and passing this off as sovereign technology. In effect, he had described his own scenario before he was caught. Choosing a ready-made model as a foundation is not prohibited in itself and is common practice in the industry. Fine-tuning existing weights is cheaper than training from scratch, and many companies do it.
What are the key details on This is not an isolated case?
A similar story already happened in September. Spain's Multiverse Computing introduced Quasar 438B as a leading European model. Only the day after the announcement did a caveat appear in a technical blog post: it is an optimized version of GLM-5.2, a Chinese model. Two cases in a row show a pattern. A "European model" on the market can mean three different things: trained from scratch in Europe, fine-tuned in Europe on someone else's weights, or simply hosted in a European data center.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.