Five months after dropping OpenAI for Mistral, Ecosia is changing AI provider again, and this time the replacement isn’t a European model. In an interview with POLITICO published on 6 October, founder and CEO Christian Kroll said the Berlin search engine is dropping Mistral in favour of open-weight models, including Chinese ones. He didn’t soften it: “We are disappointed with the quality of Mistral.” He also said the French lab’s models are now “a year behind” the competition.
The new setup runs through Melious, a German platform that hosts open-weight models on European infrastructure. POLITICO says that includes the Chinese models Qwen, GLM and Kimi. Kroll’s pitch is about cost as much as quality: “We’ve roughly cut our costs in half while improving quality and performance.”
Why Mistral lost the contract
Kroll gave POLITICO three reasons, and only one was about benchmarks.
- Reliability. Ecosia had recurring technical problems, including overloaded servers. “We were simply too large a customer for Mistral,” Kroll said.
- Climate fit. Ecosia has concerns about the large share of nuclear power in France’s energy mix, measured against its environmental goals.
- Sovereignty. “The fact that Mistral relies on investment from international investors is, in our view, not truly sovereign,” Kroll said.
Mistral didn’t respond to POLITICO. Its co-founder and chief scientist Guillaume Lample did respond at a press conference on Monday, inviting Ecosia to test the new model “as soon as possible” and offering early access. That new model is Large 4, which Mistral released on Tuesday.
This is quite a reversal. In May, Ecosia’s own blog announced it was “now running our AI almost entirely on a European model provider”. It said that gave it “a potential pathway to self-hosting our AI in the future”. Ecosia is now saying the European part that matters is the infrastructure, not the lab that trained the weights. Read alongside that third reason, sovereignty has quietly been redefined: what counts is who controls the servers your queries run on, not who built the model.
The help centre hasn’t caught up
As of this morning, Ecosia’s help page on generative AI (last updated 2 October) still says AI Chat is “powered by Mistral small 4”. It also says AI Overviews are “powered by Staan through EUSP” and use “a Mistral small 3.2 model”. Its privacy section tells users that “Ecosia forwards your chat requests to our external chat provider, Mistral”. It does add that Ecosia is “continuously exploring alternative models”, and that has turned out to be true.
The Overviews detail is the part I’d watch. Overviews run on Staan via EUSP, the European index effort behind the European search results Ecosia and Qwant launched in Germany. POLITICO doesn’t say whether the switch covers Overviews, Chat or both. Until Ecosia updates that page, it isn’t clear which Ecosia answers users see are written by which model. Users deserve that clarity, and so do publishers trying to understand why they are or aren’t cited.
What it means for search marketers
Ecosia isn’t a big traffic source for most sites, but this move matters beyond its size.
- Swap the model and the citations change. Which sources an answer engine picks and how it summarises them depends partly on the model doing the summarising. If you’re tracking AI visibility on Ecosia, treat this week as a break in the data series.
- Chinese open weights are now part of European search. Qwen is already behind Alibaba’s own web-search tooling. Now it may also be summarising results for German government users. POLITICO says Ecosia is used by institutions ranging from Germany’s federal environment ministry to the UK’s NHS.
- Bias is the open question. POLITICO cites a NewsGuard investigation that found five leading Chinese-backed models repeatedly failed to correct pro-China falsehoods. Kroll argues that geopolitical distortion can be fixed technically. Rasmus Rothe of the German AI Association told POLITICO that overt censorship can largely be retrained away, but subtle bias “embedded in the training data” is much harder: “you can scratch the surface, but you won’t reach the foundation.”
My take: Ecosia is the clearest case yet of a European search brand concluding that “European” describes where the AI runs and who the model is accountable to, not where it was trained. Searchers in Europe still have to check what’s behind the label, and this move makes that harder. For SEOs, the practical point is the same as on every other answer engine. Your page gets read by whichever model was cheapest and best this quarter, and that can change faster than your content calendar. We’ll keep tracking it in Global Search Engines.