Ecosia is ending its AI partnership with Mistral and moving to open-weight models, founder and CEO Christian Kroll told POLITICO. 'We are disappointed with the quality of Mistral,' Kroll said, calling the French flagship lab's models 'a year behind' the competition. The Berlin-based search engine, which markets itself on climate action and independence from non-European technology, is used by institutions ranging from Germany's federal environment ministry to the United Kingdom's National Health Service.
The change comes only months after Ecosia replaced OpenAI with Mistral in May — a move then presented as a bet on European AI. Kroll cited recurring technical problems, including overloaded servers. 'We were simply too large a customer for Mistral,' he said. He also raised environmental and sovereignty questions: France's electricity mix is heavily nuclear, which Ecosia considers a poor fit with its climate goals, and he argued that Mistral's dependence on foreign investors — including South Korea's Samsung Electronics and US chip giant Nvidia — undercuts claims of European independence.
Ecosia is now partnering with the AI platform Melious, which hosts open-source and open-weight models including the Chinese systems Qwen, GLM and Kimi. 'We've roughly cut our costs in half while improving quality and performance,' Kroll said. Politically, the choice fits a shift in European thinking: Kroll argues that Europe, having missed the first phase of the generative AI boom, can build on increasingly capable open-weight models instead of spending billions on frontier training runs.
The switch also illustrates the awkward politics of 'sovereign AI' in Europe. Mistral is the continent's clear frontrunner and responded through chief scientist Guillaume Lample, who said he would encourage Ecosia to test the company's new model as soon as possible, offering early access, and stressed that it takes computing power and time to build great models. Mistral released its new Large 4 model on Tuesday.
Relying on Chinese models brings its own risk. Western researchers — including a NewsGuard investigation — have found that leading Chinese-backed models repeatedly fail to correct pro-China falsehoods. Kroll argues geopolitical distortions can be fixed with technical measures, unlike quality shortfalls. That is disputed: Rasmus Rothe of the German AI Association said overt censorship can largely be retrained away, but subtle bias 'is embedded in the training data and in what the model takes for granted'.
