Le Chonk and Kolibri: Europe’s Answer to OpenAI, Built in Europe on Nvidia Chips

London, 8 October 2026 — EBM Newsdesk Analysis — By Katie Winearls

In one week, Europe’s two main AI model developers have both released their most ambitious work yet. On Saturday, Germany’s Aleph Alpha launched Kolibri, a model built for government and industry. On Tuesday, Mistral launched Mistral Large 4, nicknamed “Le Chonk”, a model with roughly one trillion parameters that it says rivals the best open models from China. Both are being sold as sovereign AI: built in Europe, trained in Europe and under European control.

For years the argument has been that Europe makes the rules on AI and America builds the technology. These two releases challenge that. The question for European businesses is whether “sovereign” means anything more than a label.

Le Chonk: Mistral Goes Big

Mistral Large 4 has about one trillion parameters, 49 billion of which are active at any one time. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell chips in its own European data centres. It covers more than 160 languages, including every official EU language, and handles documents, charts and images. A public preview is already available, and the full open weights are due later this month after security testing.

Mistral’s claims are bold. It says the model beats GPT-6-Astra on one visual benchmark and solved 93% of the Cybench hacking exercises, a test where US rivals scored close to zero, mostly because they refused parts of the task. That result says as much about safety policy as about capability, and buyers should treat it carefully. Chief executive Arthur Mensch’s message was clear, though: the idea that Europe “cannot compete” is “not true”.

Mistral has the money to back that up. Its €21bn valuation after September’s €3bn round makes it one of Europe’s most valuable private technology companies.

Kolibri: Aleph Alpha Goes Specialist

Aleph Alpha has taken a different route. Kolibri is a smaller, specialist model with 78 billion parameters, of which only about 3 billion are active at a time. That makes it cheap to run on a company’s own servers. Nearly a quarter of its training data is in German, and it is designed for public administration and regulated industries. It is trained to say it doesn’t know when its sources don’t support an answer, which matters a great deal in a government department.

Its strongest selling point is compliance. Aleph Alpha screened its training data against a blocklist of more than 4.5 million websites, including the European Commission’s piracy watch list, and has published a technical report covering copyright and the EU AI Act. For a German ministry or bank worried about where its data goes, that audit trail is the main reason to buy.

The Problem With “Sovereign”

There’s an awkward detail. In April, Aleph Alpha agreed to combine with Canada’s Cohere, and the two signed a definitive agreement last month. The combined company will operate as Cohere, with headquarters in both Canada and Germany. Germany’s flagship sovereign AI model is being released just as its developer becomes part of a Canadian company.

Both models also depend on American hardware. Le Chonk was trained on Nvidia chips. European sovereignty currently means European data centres and European law, not European silicon. While US chipmakers set the price and supply of AI computing, Europe’s independence has limits.

Why It Still Matters

None of this makes the releases unimportant. European governments and regulated industries want AI they can audit, run on their own servers and hold to account under European law. Two credible options now exist: one large general-purpose model and one smaller specialist. That gives Europe some bargaining power with US providers it didn’t have a year ago.

The cyber results also matter beyond the usual benchmark race. As attackers use AI to find and exploit weaknesses faster, European defenders want models they control that will actually help with security testing.

The Call

Mistral has shown that Europe can build a frontier-scale model. Aleph Alpha has shown that a smaller, fully documented model has a market among cautious public bodies. But true sovereignty needs three things: models, money and chips. Europe now has the first and is getting the second. Without the third, and with Aleph Alpha becoming part of Cohere, “sovereign AI” is still more an aim than a fact. European governments should give these models real contracts, because that’s how the money follows.

Related Analysis

• “Investment Advisers are Either Going to Be on the Crest of This AI Wave, or They are Going to Drown Under It”: Ray Dalio
• Inside the Criminal Networks Blurring the Lines Between Cybercrime, Crypto and Financial Fraud
• Rearming Europe: Industrial Cooperation Must Rise to the Challenge of Meeting Urgent Needs