Ecosia drops Mistral for Chinese models hosted in Europe

"We serve those as well"

On October 8, Mistral CEO Arthur Mensch replied on X to a POLITICO post. The outlet had just reported that Ecosia, the Berlin search engine that funds tree planting, is dropping Mistral for open models, Chinese ones included. His answer ran two sentences: "We serve those as well! Sorry if using nuclear energy is a real blocker."

The second sentence is a jab. The first deserves more attention, because you can check it. On August 11, Mistral announced that its platform would host open models from other developers, "starting with Z.ai's GLM-5.2."

Z.ai is the export name of Zhipu, a Chinese lab. Its version 5.3 now sits in Mistral's own model catalog. So the story of a European champion dumped for China starts with a problem: the champion serves Chinese models too.

What Ecosia has against Mistral

It starts with an interview in POLITICO, published October 6. Christian Kroll, Ecosia's founder and CEO, put it bluntly: "We are disappointed with the quality of Mistral." He called its models "a year behind" the competition. Ecosia had swapped OpenAI for Mistral in May. Since then, he said, technical problems kept piling up, overloaded servers among them: "We were simply too large a customer for Mistral."

Then comes nuclear power, and here you have to read POLITICO closely. The outlet writes that Ecosia questions whether Mistral fits its environmental goals, given how much of French electricity comes from nuclear. That's POLITICO's wording. The grievance is attributed to Ecosia, and the article contains no Kroll quote on the subject. It's still the point Mensch chose to answer.

The last complaint is a direct quote. That Mistral relies on international investors is, in Kroll's words, "not truly sovereign." The full sentence: "The fact that Mistral relies on investment from international investors is, in our view, not truly sovereign." POLITICO points out that Mistral has raised billions of euros from foreign backers, Samsung and Nvidia among them.

Mistral did not respond to POLITICO's request for comment. But its chief scientist, Guillaume Lample, asked at a press conference the day before, held out a hand. He said he would "encourage" Ecosia to test the company's new model as soon as possible, and offered "early access today."

Melious, the middleman nobody explains

To replace Mistral, Ecosia teamed up with Melious. The name appears in every write-up, and none of them explains it.

Melious AI GmbH is a German company based in the University of Saarland's start-up center in Saarbrücken. It sells one entry point to more than sixty open-weight models: Alibaba's Qwen, Zhipu's GLM, Moonshot's Kimi, DeepSeek, Google's Gemma, and also Mistral Small 4, which is on display on its website.

Melious doesn't run those models itself. Its public subprocessor list names about ten hosting providers in Europe, which execute the requests and take over when one of them goes down. Think of a booking platform rather than a hotel. Two of the rooms on offer are French: Scaleway in Paris and OVHcloud in Roubaix. Nothing says which ones will handle Ecosia's traffic.

Seen that way, Ecosia is changing its purchasing model more than its country. One supplier whose servers were buckling gets replaced by an intermediary that spreads the load. "Too large a customer" starts to mean something else.

And where do users' searches go?

That's the question every user asks on reading "Chinese models." The answer hinges on a distinction people often blur.

An open-weight model is a published recipe. Alibaba or Zhipu trains it, then puts the file online. Anyone can download it and run it on their own machines. When a European host runs Qwen, your question doesn't go back to Alibaba. It stays in the kitchen of whoever downloaded the recipe.

Calling a Chinese lab's own API is different: your question goes straight to them. In September we saw what that can look like. According to an Anthropic report, Moonshot was relaying requests from Kimi customers to Claude without their knowledge (we covered it here).

According to Melious's standard contract, Ecosia's setup falls under the first case. That contract provides for processing in Germany, the EU or the European Economic Area, requires the customer's written consent for any transfer outside that zone, and states that message content is not logged. Two caveats. This is Melious's generic contract: Ecosia's isn't public. And on October 2, Ecosia's help page still said it "forwards your chat requests to our external chat provider, Mistral."

What the file still carries

If the data doesn't travel, something else comes along with the recipe: what the model learned. POLITICO recalls a NewsGuard investigation finding that leading Chinese models repeatedly fail to rebut false claims that favor Beijing.

For Kroll, that's not a deal-breaker: such distortions can be fixed technically, unlike poor quality. Rasmus Rothe, head of the German AI Association, agrees only halfway. Visible censorship, the refusals to answer, largely fades with targeted retraining. Subtle bias holds, because it's baked into the training data: "With a few hundred example questions, you can scratch the surface, but you won't reach the foundation."

Price, the argument nobody puts a number on

That leaves the most down-to-earth motive. With Melious, Kroll says, "We've roughly cut our costs in half while improving quality and performance." Which costs, over what period, compared with what? The interview doesn't say. It's the claim of a customer happy with its new supplier, and it can't be verified from the outside.

Ecosia isn't the only one talking money. Qwant, its partner on the European search index, still powers its flash answers with Mistral, but told Next, a French tech outlet, that it is evaluating alternatives because of "Mistral's recent pricing changes." Qwant didn't say which ones.

The public price of Mistral Small 4, which runs Ecosia's chat according to its help page, is now $0.15 per million input tokens and $0.60 per million output tokens. The third-party tracker UsagePricing had it at $0.10 and $0.30 at the end of May. A big customer doesn't necessarily pay the list price.

So two European buyers are talking about cost and quality, two months after Mistral opened its own storefront to a Chinese model. Kroll isn't giving up on sovereignty. He even faults Mistral for falling short of it. But the word has moved. It used to mean the lab that trains the model. Now it means the server that runs it. And on that ground, Mistral and Melious offer the same thing.