Ecosia drops Mistral for Chinese open weights on EU-hosted Melious

Ecosia, the Berlin-based search engine used by Germany's federal environment ministry and the UK's National Health Service, is dropping Mistral AI as its model supplier in favor of open-weight models, including Chinese families Qwen, GLM, and Kimi, served through EU-hosted infrastructure. CEO Christian Kroll says costs roughly halved while quality improved, and calls Mistral's models "a year behind" the competition.

Key facts

• Ecosia replaced OpenAI with Mistral AI in May 2026 and is now switching to open-weight models hosted by Melious, a German platform serving models on servers in eight EU countries (POLITICO Europe).
• Kroll's claim: "We've roughly cut our costs in half while improving quality and performance", a customer self-report rather than an audited figure (POLITICO Europe).
• Mistral chief scientist Guillaume Lample responded on October 6: "I would encourage them to test our new model as soon as possible. We can even give them early access today if they want" (POLITICO Europe).
• A NewsGuard investigation found the five leading Chinese-backed AI models repeatedly failed to correct pro-China falsehoods (NewsGuard).
• Mistral released Large 4 on October 6, the same week Ecosia's switch became public (POLITICO Europe; open weights are promised by the end of October, per our Large 4 preview coverage).

What happened

Kroll told POLITICO Europe that Ecosia's relationship with Mistral AI, which began in May 2026 when the search engine replaced OpenAI, ran into both quality and capacity problems. "We are disappointed with the quality of Mistral," he said, describing the French lab's models as "a year behind" the competition. He also said Ecosia hit recurring technical problems, including overloaded servers: "We were simply too large a customer for Mistral."

The replacement is Melious, a German platform that hosts open-weight models on servers spread across eight EU countries. Its catalog includes the Chinese open-weight families Qwen (Alibaba), GLM (Z.ai), and Kimi (Moonshot AI). POLITICO could not independently confirm the details of the hosting arrangement, and Ecosia has not published which specific models route which traffic. What is public is Kroll's characterization of the result: "We've roughly cut our costs in half while improving quality and performance." That is the customer's own accounting. What would verify it is a published price sheet from Melious and an independent quality comparison. Neither exists yet.

Kroll framed the move as principled as well as practical. He raised questions about whether Mistral guarantees European independence, pointing to the lab's funding: "The fact that Mistral relies on investment from international investors is, in our view, not truly sovereign." Mistral has raised billions of euros from abroad, including South Korea's Samsung Electronics and Nvidia. He also questioned whether France's nuclear-heavy energy mix aligns with Ecosia's climate positioning.

Mistral AI did not respond to POLITICO's request for comment. But its chief scientist and co-founder Guillaume Lample, speaking at a press conference on October 6, made a public pitch for the customer to come back: "I would encourage them to test our new model as soon as possible. We can even give them early access today if they want." The timing stings. Mistral released Large 4 that same Tuesday, and the model's open weights are promised by the end of October.

Why it matters

This is the most concrete EU-sovereignty data point since European buyers started asking whether "European AI" means models trained in Europe or models usable under European jurisdiction. A European institution-serving buyer with a European-first brand looked at the European flagship lab and chose Chinese open weights on EU-hosted infrastructure instead, on cost, quality, and reliability grounds rather than political ones. That inverts the usual sovereignty argument, in which open weights are the hedge and the domestic lab is the preference.

It also matters because of who Ecosia serves. Germany's federal environment ministry and the UK's National Health Service are institutional users with procurement standards, not consumers experimenting with a chatbot. When a buyer at that level says Chinese open-weight models beat the European flagship on quality and price, procurement officers across the EU will read the interview.

For builders, the structural signal is familiar from the enterprise side of the Atlantic: AT&T told the Financial Times it runs about 40% of its AI workloads on open models, targeting 70% (see our AT&T open-weights coverage). Ecosia's switch is the European counterpart at consumer-product scale. The pattern in both cases: buyers are discovering that open weights plus competent hosting is now competitive with renting a lab's API, and the lab's brand is not a moat.

Background

Ecosia built its brand on independence from non-European tech, first from Google, then from OpenAI, which it replaced with Mistral AI in May 2026. That the replacement lasted roughly five months before Ecosia moved again, this time to open weights, says something about how fast the middle of the market is repricing. Our Mistral Large 4 preview coverage noted the lab's open-weights bet: 1T parameters in a mixture-of-experts architecture, with weights promised by the end of October. Ecosia's departure landed the same week.

The open-weight route has a known objection, and it deserves to be stated plainly. NewsGuard's investigation found the five leading Chinese-backed AI models repeatedly failed to correct pro-China falsehoods, and a US advisory body has warned that China's open-source dominance threatens the American AI lead (Reuters). Kroll's position is that these biases are not a dealbreaker because "geopolitical distortions can be addressed through technical measures," unlike shortcomings in raw model quality.

That is disputed. "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 harder, he argues, "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." Both positions can be true at once: retraining fixes refusals, and nothing yet fixes priors. Buyers evaluating Chinese open weights for EU-facing products should test for both, and our EU AI Act enforcement coverage tracks the transparency obligations that apply to whatever model a European product ships. For the underlying licensing mechanics, see our explainer on what open-weight models are.

What's next

Mistral Large 4's open weights are promised by the end of October, with a countdown on the Hugging Face repo pointing to an October 31 ETA. If Lample's "early access today" offer gets taken up, watch for an Ecosia pilot of Large 4, the cheapest possible reversal of this story. On the other side, Melious remains the weakly documented link in Ecosia's chain: no independent verification of its EU-hosting footprint or its hosted-model list has been published, and any builder evaluating it should demand both. The measurable things to watch next: Melious's published pricing, an independent quality eval of the Ecosia deployment, and whether other EU institutional buyers follow.