Ecosia, the Berlin-based search engine used by institutions ranging from Germany's federal environment ministry to the UK's National Health Service, is dropping Mistral, its AI provider since May, in favor of open-source and open-weight models, CEO Christian Kroll told POLITICO. The models will be hosted by the AI platform Melious, which offers a range of open models including Qwen, GLM and Kimi. "We are disappointed with the quality of Mistral," Kroll said, describing the French lab's models as "a year behind" the competition. In May, Ecosia had replaced OpenAI with Mistral.
Why Ecosia left
Quality comes first. Capacity is second: Ecosia saw recurring technical problems, including overloaded servers. "We were simply too large a customer for Mistral," Kroll said. Mistral's compute business is covered in Mistral Stops Selling Compute and Starts Selling Trust.
Fit is the rest. Ecosia has built its brand on climate action and independence from non-European tech, and it questions whether Mistral matches its environmental goals given the large share of nuclear power in France's energy mix. It also doubts the sovereignty claim: "The fact that Mistral relies on investment from international investors is, in our view, not truly sovereign," Kroll said. Mistral has raised billions of euros from abroad, including from South Korea's Samsung Electronics and Nvidia, as covered in Mistral's record Series D.
On the numbers, Kroll said: "We've roughly cut our costs in half while improving quality and performance." He sees open-weight models, championed by China, as a major opportunity: Europe missed the first phase of the generative AI boom, but can build on increasingly powerful open models without spending billions to train frontier systems. That mirrors an approach backed by EU leaders and advisors as they reorient the bloc's AI strategy.
Kroll's clarification
Speaking to EU-Startups, Kroll added that the decision is not a move toward Chinese technology but a shift toward more affordable, energy-efficient open AI models. "We'll probably be using Melious AI in future," he said. "Unlike Mistral, this is a European company that offers high-quality, affordable AI inference powered by green electricity."
Mistral's response
Mistral did not respond to POLITICO's request for comment before Guillaume Lample, its chief scientist and co-founder, spoke at a press conference on Monday. "I would encourage them to test our new model as soon as possible," he said. "We can even give them early access today if they want." Building great models takes time, computing power and science, he said, and Mistral can give customers the weights of its large and small models and customize them together.
CEO Arthur Mensch replied on X to the POLITICO post: "We serve those as well! Sorry if using nuclear energy is a real blocker."
Mistral released Large 4 on Tuesday, a launch covered in Mistral Large 4 'Le Chonk', and National Rally President Jordan Bardella said it "puts France (and Europe) back in the race for artificial intelligence." With Mistral the clear frontrunner for AI in Europe, Ecosia's switch raises questions about the bloc's top company and shows the scale of the challenge of building a competitive, sovereign model on European soil, a theme of France's economy minister's warning that European AI cannot be just Mistral.
The China question
Amid strong US national security concerns about Chinese models, Western research organizations have also questioned their effectiveness given censorship on politically sensitive topics. A recent NewsGuard investigation found that the five leading Chinese-backed AI models repeatedly failed to correct pro-China falsehoods.
Kroll says such biases are not a dealbreaker. Geopolitical distortions, he argues, can be addressed through technical measures, unlike shortcomings in model quality. Rasmus Rothe, executive director of the German AI Association, says overt censorship, meaning refusals or evasive answers, "can be largely eliminated through targeted retraining" and already is in practice. Subtle bias is harder because it is embedded in the training data and in what the model takes for granted. "With a few hundred example questions, you can scratch the surface, but you won't reach the foundation," Rothe said.