“We are disappointed with the quality of Mistral,” Ecosia founder and CEO told POLITICO.
AI generated Text-to-speech
BERLIN — A Berlin-based search engine used by government services is dropping its French AI partner, Mistral, in favor of open-source models, including those from China, its CEO told POLITICO.
“We are disappointed with the quality of Mistral,” Ecosia founder and CEO Christian Kroll said in an interview, stating that the models from France’s flagship AI lab are now “a year behind” the competition.
Ecosia — a German web browser and search engine that has built its brand on support for climate action and independence from non-European tech — is used by institutions ranging from Germany’s federal environment ministry to the U.K.’s National Health Service.
In May, it replaced the American giant OpenAI with Mistral as its AI provider. But the company experienced recurring technical problems, including overloaded servers, Kroll said. “We were simply too large a customer for Mistral,” Kroll said
Ecosia also has concerns about whether Mistral aligns with its environmental goals, given the large share of nuclear power in France’s energy mix and questions whether it guarantees European independence. “The fact that Mistral relies on investment from international investors is, in our view, not truly sovereign,” Kroll said.
Mistral did not respond to a request to comment for this article. However, Guillaume Lample, chief scientist and co-founder of Mistral, told POLITICO at a press conference Monday: “I would encourage them to test our new model as soon as possible. We can even give them early access today if they want.”
“And really, the idea here is that it takes time to build great models. It takes computing power, it takes science, and we truly believe we’re growing exponentially. So I encourage them to ride that exponential growth with us, but also to take advantage of our comprehensive customization approach. We can provide customers with the weights of our various models — the large ones, but also the smaller ones — and we can customize them together with them,” Lample said.
Mistral on Tuesday released a new model, Large 4, which far-right National Rally President Jordan Bardella said“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 illustrates the scale of the challenge of building a competitive, sovereign AI model on European soil. Mistral has raised billions of euros from investors abroad, including South Korea’s Samsung Electronics and American giant Nvidia.
Kroll said he sees open-weight models, championed by China, as a huge opportunity for Europe. While the continent missed the first phase of the generative AI boom, the German entrepreneur said Europe can build on increasingly powerful open-source models without investing billions in training frontier systems.
That mirrors an approach being backed by EU leaders and advisors as they seek to reorient the bloc’s AI strategy.
To that end, Ecosia is partnering with the AI platform Melious, which hosts a range of open-source and open-weight models, including Chinese models Qwen, GLM, and Kimi. “We've roughly cut our costs in half while improving quality and performance,“ Kroll said.
Amid strong U.S. concerns about the national security implications of Chinese models, Western research organizations have also raised concerns about the models’ effectiveness given their censorship on politically sensitive topics. A recent NewsGuard investigation found the five leading Chinese-backed AI models repeatedly failed to correct pro-China falsehoods.
But Kroll said such biases are not a dealbreaker, arguing that geopolitical distortions can be addressed through technical measures — unlike shortcomings in model quality.
That’s disputed by some experts. “Overt censorship — that is, refusals to answer or evasive answers ... can be largely eliminated through targeted retraining. This is well-documented technically and is already being done in practice,” said Rasmus Rothe, executive director of the German AI Association.
Subtle bias is more difficult to address, “because it’s 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.