Ecosia drops Mistral for Chinese open-weight models — but keeps the inference in Europe

Ecosia has stopped using Mistral AI to generate its AI search answers, five months after adopting it. Founder and CEO Christian Kroll told Politico the company was disappointed with the quality, describing Mistral as roughly a year behind competitors and its servers as regularly overloaded — Ecosia, he said, had simply become too large a customer for the infrastructure available. The switch was reported on 8 October 2026.

In its place, Ecosia is running open-weight models: Qwen from Alibaba, GLM from Zhipu AI, and Kimi from Moonshot AI. Kroll says the change roughly halved operating costs while improving quality and performance at the same time.

The detail most coverage has skipped is where those models actually run. Ecosia’s partner is Melious, operated by Melious AI GmbH in Germany, which serves open-weight models across eleven infrastructure providers in eight EU countries and states that data never leaves European jurisdiction. So the model weights are Chinese; the inference is not. For a reader choosing tools on sovereignty grounds, those are different questions, and conflating them gets the answer wrong in both directions.

Update, 11 October. Kroll has pushed back on the framing in an interview with EU-Startups, saying the decision is about open models rather than their country of origin: “We’ll probably be using Melious.ai in future. Unlike Mistral, this is a European company that offers high-quality, affordable AI inference powered by green electricity.” Note the tense — on his own account the move to Melious is a direction rather than a finished migration. He also says Ecosia can route to a different model on sensitive subjects such as Taiwan, and is working to strip political bias from its answers. Melious co-founder Simon Jakobs put the same point in infrastructure terms: whether a model was built in Europe, the United States or China, “it’s simply a set of weights on our servers”, outside the reach of the US CLOUD Act, with customer data never retained or used for training.

Kroll’s other two reasons are worth separating as well. He pointed to the nuclear-heavy French electricity mix as a poor fit for a company that sells itself on environmental grounds — a judgement about carbon accounting rather than about software. And he questioned how European Mistral still is, given investment from Nvidia and Samsung. That last point cuts in an uncomfortable direction, because it is the same test this site applies to every entry, and Mistral passes it: the company remains French-headquartered with European-hosted inference, and foreign minority investors do not transfer control. Mistral’s September 2026 Series D raised roughly EUR 3 billion led by Samsung Electronics at a valuation above EUR 21 billion, with Andreessen Horowitz, Nvidia and Salesforce Ventures alongside ASML and Bpifrance — a cap table with substantial American money in it, and still no American owner.

The contrast Kroll draws with his own company is real, and it is structural rather than rhetorical. Ecosia adopted a steward-ownership model through the Purpose Foundation in 2018: the company cannot be sold, and nobody inside or outside it holds economic rights to it. That is the one arrangement a funding round cannot undo.

The timing is pointed. Mistral shipped Mistral Large 4 on 6 October, two days before the news broke, so the “year behind” judgement describes the models Ecosia was actually served rather than the one that had just launched. Vibe and Vibe for code both run on the current generation.

It is also only half of Ecosia’s year. In May 2026, Ecosia and Qwant began serving results from EUSP, the independent search index the two companies built together to escape their dependence on Bing and Google. Ecosia has moved toward European infrastructure on the part that is hardest to replace, and away from a European supplier on the part that is easiest to swap. Both decisions were made on the merits, and that is roughly what a maturing market should look like.