Ecosia, the Berlin-based search engine used by institutions from Germany's federal environment ministry to the UK's National Health Service, is replacing Mistral as the supplier of its artificial intelligence models, according to an interview with chief executive Christian Kroll that POLITICO published on October 6, 2026. Kroll described the French company's models as "a year behind" the competition.

In Short

Ecosia, a Berlin search engine, is stopping its use of AI models from the French company Mistral and switching to models whose files anyone can download, some of them made in China, because its boss says Mistral's were falling behind and its servers kept getting overloaded. It matters because Mistral is Europe's best-known AI company, so losing a customer that uses a lot of capacity is a visible setback for the idea of a European-built AI stack. The practical change is the supplier: Ecosia is working with a hosting platform called Melious, and its chief executive says costs fell by about half.

A customer leaves Europe's best-known AI lab

In May, according to POLITICO, Ecosia replaced the American company OpenAI with Mistral as its AI provider. The arrangement ran into difficulty. Kroll told the publication that the company experienced recurring technical problems, among them overloaded servers, and put the cause in blunt terms: "We were simply too large a customer for Mistral." Quality was the second complaint. "We are disappointed with the quality of Mistral," he said in the interview.

Two further objections have less to do with engineering. Ecosia has built its brand, POLITICO notes, on support for climate action and independence from non-European tech, and the company has concerns about whether Mistral fits its environmental goals given the large share of nuclear power in France's energy mix. The article does not elaborate on how that energy mix bears on Ecosia's criteria. On sovereignty, Kroll was direct: "The fact that Mistral relies on investment from international investors is, in our view, not truly sovereign." Mistral has raised billions of euros from investors abroad, according to POLITICO, including South Korea's Samsung Electronics and the American company Nvidia.

The supplier arrangement has therefore changed twice since May: from OpenAI to Mistral, and now to a hosted mix of open-weight models.

Open-weight models and the Melious partnership

Kroll sees open-weight models, championed by China, as a major opportunity for Europe, according to POLITICO. The continent missed the first phase of the generative AI boom, the German entrepreneur said, but it can build on increasingly powerful open-source models without spending billions of euros training frontier systems. POLITICO adds that the approach mirrors one being backed by EU leaders and advisers as they seek to reorient the bloc's AI strategy.

An open-weight model is one whose trained parameters are published for others to download and run; the training data and code often stay private. The label is narrower than open source in the strict sense, although POLITICO uses the two terms side by side. Mistral, too, distributes weights, so the dispute concerns which models perform better and who stands behind them rather than whether they are open.

To put the strategy into practice, Ecosia is partnering with Melious, an AI platform that hosts a range of open-source and open-weight models. Among them are the Chinese models Qwen, GLM and Kimi, developed by Alibaba, Zhipu AI and Moonshot AI respectively. "We've roughly cut our costs in half while improving quality and performance," Kroll said. The figure is his own, and POLITICO prints no baseline, benchmark or invoice data against which to test it.

Mistral's reply and the Large 4 release

Mistral did not respond to POLITICO's request for comment. Its chief scientist and co-founder, Guillaume Lample, did address the matter at a press conference on Monday, October 5, according to POLITICO. "I would encourage them to test our new model as soon as possible," he said. Progress takes time, Lample argued: "it takes time to build great models," and he added that "we truly believe we're growing exponentially." He also pointed to what he called a "comprehensive customization approach", under which Mistral can supply customers with the weights of its large and smaller models and adapt them jointly with those customers.

The model in question, Large 4, was released on Tuesday, October 6, according to POLITICO. National Rally President Jordan Bardella, whom POLITICO describes as far-right, said it "puts France (and Europe) back in the race for artificial intelligence."

Mistral's own announcement gives the technical outline. Large 4 is a natively multimodal model with 1 trillion parameters, of which 52 billion are active - the portion of the network used to process any single token. A public preview is available through Mistral Studio, and the company says it will release the weights by the end of October. According to Mistral, the model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own datacentres in Europe.

The announcement also reports a blind human coding evaluation run with Surge AI. In it, according to Mistral, the Large 4 preview ranked second among five models with a score of 3.74, behind only Claude Opus 5 at 4.22 and ahead of Kimi K3 (3.59), GLM-5.3 (3.60) and GLM-5.2 (3.40). The test is vendor-reported, limited to coding and run on a preview build. Kimi and GLM are among the model families Ecosia is adopting, though POLITICO does not say which versions.

POLITICO frames the episode in wider terms. With Mistral the clear frontrunner for AI in Europe, the outlet writes, Ecosia's switch raises questions about the bloc's top company and illustrates the scale of the challenge of building a competitive, sovereign AI model on European soil.

Bias, censorship and the NewsGuard findings

Choosing Chinese models carries its own objections. POLITICO reports strong US concerns about the national security implications of such systems, and says Western research organisations have questioned their effectiveness given censorship on politically sensitive topics. It cites "a recent NewsGuard investigation" finding that the five leading Chinese-backed AI models repeatedly failed to correct pro-China falsehoods, without naming a date for the report.

NewsGuard's own site lists two relevant publications. A July 25, 2025 audit covered five Chinese models - Baidu's Ernie, DeepSeek, MiniMax, Alibaba's Qwen and Tencent's Yuanbao - and found a 60 percent failure rate on English-language prompts across ten false claims. A report dated August 13, 2026 covered seven Chinese chatbots, which failed to debunk pro-China false claims 53 percent of the time, against 24 percent for 10 Western chatbots used as a benchmark. The five-model count in POLITICO matches the earlier audit, while the more recent report covers seven. POLITICO does not specify which one it means.

Kroll said such biases are not a dealbreaker. Geopolitical distortions, he argued, can be addressed through technical measures, unlike shortcomings in model quality. Some experts dispute that. Rasmus Rothe, executive director of the German AI Association, told POLITICO that overt censorship, meaning refusals to answer or evasive answers, "can be largely eliminated through targeted retraining". Subtle bias is harder, Rothe said, "because it's embedded in the training data and in what the model takes for granted." His summary of the limits of the fix: "With a few hundred example questions, you can scratch the surface, but you won't reach the foundation."

Where the move sits in the search market

Ecosia is not only a consumer of AI models. On August 7, 2025 it began serving search results from its own European index, the first deployment of the European Search Perspective, a 50-50 joint venture with the French search engine Qwant. The index runs on Staan, a system that handles crawling, indexing, ranking and serving, and it offers an API aimed at AI applications. The company aimed to serve 50 percent of French queries through the index by the end of 2025. The model supplier sits on a separate layer of that stack, and POLITICO does not say which Ecosia features depend on the models.

That separation matters for a reason that reaches beyond Ecosia. AI-generated answers in search have become a regulatory subject in Europe: on December 9, 2025 the European Commission opened a formal antitrust investigation into whether Google abused a dominant position by using publisher content and YouTube videos for AI purposes without fair compensation or workable opt-out options. The coverage cited research linking AI summaries to click-through rate declines of between 34.5 percent and 54.6 percent. For publishers and advertisers, the practical interest lies in the layer that decides what an AI answer cites and what it leaves out, and POLITICO's article does not say how Ecosia's models handle sources or advertising.

The regulation of the models themselves is a third thread. Obligations on providers of general-purpose AI models under the EU AI Act fall within the remit of the AI Office, which has been able to exercise enforcement powers since August 2, 2026. A draft implementing regulation dated March 12, 2026 lists APIs, source code and model weights among the ways the Commission may require access to a model, and the Act allows fines of up to 3 percent of global annual turnover or 15 million euros, whichever is higher. POLITICO's article does not discuss how those obligations apply to Chinese-developed weights served by a European platform.

What the source does not establish

Several points remain open in the reporting. POLITICO does not say when the switch will be complete or whether Mistral models will be removed entirely, since it describes Ecosia as "dropping" the partner. It does not say whether Ecosia has taken up Lample's offer to test Large 4, where Melious runs its infrastructure, or which Ecosia products run on the models. Lample's remarks, as reported, did not address Kroll's account of overloaded servers.

Timeline

Summary

  • Who: Ecosia, the Berlin-based search engine led by founder and chief executive Christian Kroll; Mistral, the French AI lab; and Melious, the AI platform hosting the open-weight models Ecosia is adopting. POLITICO reporters Anouk Schlung and Larissa Kögl conducted the interview.
  • What: Ecosia is dropping Mistral as its AI model supplier in favour of open-source and open-weight models, including the Chinese models Qwen, GLM and Kimi, citing quality, recurring technical problems, environmental and sovereignty concerns, and, according to Kroll, costs cut roughly in half.
  • When: POLITICO published the interview on October 6, 2026, the same day Mistral released Large 4. Ecosia had replaced OpenAI with Mistral in May.
  • Where: Berlin, Germany, with Mistral based in France and the models hosted through Melious. The institutions using Ecosia span Germany and the United Kingdom.
  • Why: According to Kroll, Mistral's models are a year behind the competition, its servers could not handle Ecosia's volume, and its reliance on international investors falls short of his definition of sovereign. Open-weight models, in his view, let Europe build on existing systems without paying to train frontier ones.