Search engine Ecosia swaps Mistral for Chinese A.I. Europe's push for A.I. sovereignty has a problem, and it comes from Berlin: Ecosia, the search engine that is also used by public authorities, has dropped Mistral as its A.I. supplier and is turning to open models instead, including Chinese ones such as Alibaba's Qwen, Z.ai's GLM and Moonshot AI's Kimi. That is according to Politico. The timing is awkward for Mistral, which unveiled Large 4, its most powerful model yet, this very week. Even that model falls short of the best Chinese open-weight models. Disappointed by the quality. Ecosia switched from OpenAI to Mistral only in May, as a statement of European independence. According to Ecosia's help page, A.I. summaries in search most recently ran on Mistral Small 3.2 and the A.I. chat on Mistral Small 4. Now comes the reversal: Christian Kroll, Ecosia's chief executive, is disappointed with the quality of Mistral's models, which he says lag about a year behind the competition, Politico reports. According to reports, Mistral's servers were also frequently overloaded. Mr. Kroll also questions Mistral's status as a sovereign provider, because the French company depends on money from international investors. Twice a week for free - never miss a story! Ecosia now gets its models through Melious, a German platform that runs open A.I. models on servers in eight E.U. countries. According to Melious, data never leaves Europe, and its lineup includes Z.ai's GLM 5.3, Moonshot AI's Kimi K3 and DeepSeek V4.1 Flash. Reports on the switch say it roughly halved the cost of Ecosia's A.I. services while improving performance. Mistral strikes back with Large 4. This week, Mistral introduced Large 4, a model with one trillion parameters trained in European data centers, with open weights due at the end of October. The leap over its predecessors is huge: Large 4 scores 38.4 points on the Intelligence Index from Artificial Analysis, compared with 14 points for Medium 3.5 and just 9 for Large 3. That makes it the strongest open model outside China. It still falls short of the Chinese leaders, though. In the ranking of open models, Large 4 comes in eighth, and all seven models ahead of it are from China: Price is no clear win for Mistral either. At $1.13 (about €1) per task in the index, Large 4 is cheaper than GLM-5.3 and Kimi K3 at around $2 each, but DeepSeek V4.1 Flash scores slightly higher for just $0.27 (about 24 cents). Because Mistral is still training the preview version, Large 4 could improve before the weights are released. The sovereignty paradox. The case highlights a dilemma for European A.I. policy. Anyone in Europe who wants top-tier open models currently ends up almost inevitably with Chinese developers, even if those models run on European servers. Mistral itself is betting on Chinese models: Since the summer, the company has been remaking itself into a European neocloud that sells computing capacity, model hosting and regional data control as a package. For the first time, models from other developers run on its platform, and the first one was, of all things, Z.ai's GLM 5.2. Mistral has since moved to the successor, GLM 5.3, and its model overview lists no other third-party models so far. In other words, Mistral offers its customers the very Chinese A.I. that Ecosia has now switched to. Z.ai has been on the U.S. entity list since 2025. Mistral is pushing back. Large 4 beats the Chinese models in some areas, Arthur Mensch, Mistral's chief executive, told Reuters, pointing to its cyber defense capabilities, among other things. The company also stresses that Large 4 was trained entirely in Europe and that customers can run the weights on their own servers. Aus Datenschutz-Gründen ist dieser Inhalt ausgeblendet. Die Einbettung von externen Inhalten kann in den Datenschutz-Einstellungen aktiviert werden:

Ecosia adds live-web AI Chat and an ai-free search option. Ecosia's updated AI Search documentation describes several paths through the same search service: conventional results, generated overviews and conversational answers. For SEO teams, the significance is that the query alone no longer fully describes the experience being measured. Routing, mode selection and user preferences can determine which surface presents a source. What the official documentation confirms. The Help Center page updated October 2, 2026 identifies Mistral Small 4 as the model for AI Chat and Mistral Small 3.2 for AI Overviews. Overviews summarize search information and link sources above the results. AI Chat includes Web Source Mode, which instructs it to use the live web for all queries, alongside Guided Learning and Think Longer. Ecosia also documents automatic routing of some complex questions from standard results to AI Chat. Its desktop AI-free setting disables Overviews and AI routing, removes Chat from navigation and autocomplete, and restores the traditional search box. Mobile support is described as future work. AI Chat remains a beta available to some users, so the documentation should not be read as proof that every user receives identical features. The page's update date establishes when this description was revised. It does not establish that every listed feature launched together that day. Nor does a model name reveal the full process through which a source is retrieved, selected and presented. Visibility tests need to record the route. Its editorial interpretation is that Ecosia should be tested as several distinct experiences rather than reduced to one generic "AI search" score. A useful protocol would repeat a controlled set of queries in conventional results, Overviews where available and Chat, with Web Source Mode recorded explicitly. Include an AI-free desktop session as a separate condition rather than blending it into an average. Capture the date, device, region, account state, enabled settings and whether routing occurred. Then preserve the actual result or answer before classifying brand mentions and source links. Those details help distinguish a genuine change in exposure from a change in the interface used to observe it. For example, a site appearing in ordinary results may not appear among the sources in an answer. Conversely, an answer may name a business without linking its site. Treat result position, citation presence and accurate representation as separate observations. None alone tells you how many people clicked or whether the business gained a qualified visit. Live-web sourcing does not establish freshness by itself. Web Source Mode makes the retrieval condition particularly worth testing. Using a live web source does not guarantee that the underlying page is current, that the newest information was selected or that the answer preserves an important qualification. Publishers can make dates, evidence and limitations explicit; measurement must still establish what the assistant actually used. The AI-free option creates another reason to retain conventional search analysis alongside GEO monitoring. Teams should avoid interpreting a change in chatbot exposure as a complete account of discoverability on Ecosia. The relevant question is which experiences their audience uses, and what each one shows. That approach fits NetContentSEO's research-entity recognition experiment, which records observations over time without treating one successful answer as a solved visibility problem. Ecosia's documentation gives researchers additional conditions to record. It does not supply citation-selection rules or traffic-impact measurements. The practical opportunity is a better controlled audit of where sources appear, how they are described and whether those observations repeat.