Giotto Alto: what it is, is it free, and the Kimi claims
2026-10-28 · 4 min read
Giotto Alto is the flagship reasoning model from Giotto, a Lausanne AI lab, and yes, it is free for individuals in the Giotto app and at cloud.giotto.ai. Its smaller sibling, Giotto Solo, is built to run on a single GPU so a company can host it on its own hardware.
Giotto announced both on X on October 27, calling them "the smartest European sovereign AI models on the market." The post passed 340,000 views in about a day. It also picked up a community note saying the benchmark chart uses self-reported scores and that independent leaderboards don't list Giotto yet. Two researchers then went after the launch on X, and their posts are getting as much attention as the models.
What Giotto launched
Per the company's press release, Alto runs on eight B300 GPUs and is aimed at heavy reasoning and enterprise work. Solo is pitched as the smartest model that fits on one GPU. Hosting is in Europe, with no US cloud involved. Giotto, an EPFL spinout from 2017, placed 2nd of roughly 1,500 teams in the 2025 ARC-AGI competition. The release gives no parameter count, no context window, and no base model.
How to try it
| Option | What you get | Cost |
|---|---|---|
| Web chat at cloud.giotto.ai | Alto, with an anonymous mode, no account needed | Free |
| iOS and Android apps | Alto in a phone app | Free |
| API | Model name giotto-model |
€2.10 per 1M input tokens, €10.50 per 1M output, €0.21 cached input |
| Dedicated cloud | Your own instance in Switzerland or the EU, from one GPU | Monthly rental, no public price |
| On-premises | Solo on a workstation up to Alto on server racks | Annual license priced by GPU, no public price |
The per-token prices come from Giotto's API page. The press release also mentions a flat-fee plan with unlimited use, one request at a time.
For a first test, give Alto a real problem from your business, like turning a messy job description into a clean quote, and give the same prompt to the model you already use. Compare the answers line by line. Keep anything confidential out of it for now, for reasons below.
The Kimi question
On October 28, Charlie O'Neill (@oneill_c) posted a sharp reply. He claims Alto "is trained on kimi" and says the launch arrived eight days after Giotto's CEO called open source a "trojan horse" for Europe. He also says the benchmarks are dated and saturated, with no agentic tests. To be fair to him, he says plainly that building on open weights is legitimate; his complaint is about how it was presented.
The Kimi claim rests on screenshots. The news account @AGTPinsights summed it up: some users showed Alto saying it is a Kimi model from Moonshot AI, served as "giotto-model." Be careful with that: chatbots get their own identity wrong all the time, so a self-description proves nothing about the weights. Giotto has not named its base model, and we found no public response from the company to the claim.
One thing we noticed ourselves: the API's €2.10 and €10.50 per million tokens sit very close to the $2 and $10 Moonshot charges for Kimi K3.1 Pro. Similar prices prove nothing about where a model came from, but it's an odd match.
The license line that started the fight
Alexander Doria (@Dorialexander), who builds open AI infrastructure at Pleias, quoted terms that bar users from reverse engineering the model or deriving its "parameters or architecture," and from benchmarking it for publication without Giotto's written consent. His comment: "Wonder why."
We read the EULA posted on giotto.ai on October 28. The reverse-engineering ban is there, and so is a line limiting that license to "non-commercial testing and evaluation purposes." The benchmarking clause Doria quoted was not in the version we saw. The API agreement allows business use, but it bars you from using Alto's outputs to train a competing model.
Our read
A free, capable model hosted in Europe would be useful, especially for firms with GDPR obligations. But Giotto won't name its base model, and its benchmark chart already carries a community note. Until outside testers can run their own numbers, "sovereign" is Giotto's word, not a finding. If European origin is the point for you, Mistral Large 4 is the comparison: Mistral published the parameter count and promised open weights. And what's behind a chatbot does matter. In September, Anthropic alleged that two Chinese apps were quietly passing user prompts to Claude and serving the answers as their own.
So for a small business: use the free tier on work you wouldn't mind a stranger reading, and hold off on client data until Giotto says what Alto is built on. If you're thinking about the API, read the agreement first.
Choosing which model gets which job is a decision we help Fox Valley owners make. If you want a second opinion on where AI fits in your workflow, our free process audit is a good place to start.