Recurring technical problems, overloaded servers and dissatisfaction with model quality pushed Ecosia to abandon Mistral, CEO Christian Kroll told Politico. The German search company is turning to open models, including alternatives developed in China.
Ecosia had replaced OpenAI with Mistral in May, according to the report. Its planned departure reverses that earlier provider choice and brings a pointed customer complaint against one of Europe’s most prominent AI developers.
The public record, however, does not establish a completed migration. Ecosia’s AI Search help page still names Mistral Small 3.2 for AI Overviews and Mistral Small 4 for its conversational AI Search feature.
Kroll’s account establishes the company’s dissatisfaction and intended direction. The documentation leaves unresolved which features have changed providers, which replacement models are serving users and when the transition takes effect.
Kroll’s Complaint Covers Both Models and Infrastructure
“We are disappointed with the quality of Mistral,” Kroll told Politico, describing the French company’s models as “a year behind” the competition.
The report provides no benchmark, defined set of test questions, or model versions and inference settings behind that comparison. Kroll’s assessment therefore cannot establish that every Mistral model trails every alternative, or quantify how far behind any particular model might be.
He also described recurring technical problems, including overloaded servers:
“We were simply too large a customer for Mistral.”
Model quality and service capacity both matter for a consumer search product, but assessing them requires different evidence. Answer quality concerns whether a system produces useful, accurate responses. Server capacity concerns whether the provider can deliver those responses reliably under demand.
A model could perform well in an evaluation while its hosted service struggles with a customer’s traffic. A reliable service can also deliver answers that a customer considers inadequate. Kroll raises both issues without providing enough operational detail to determine their relative contribution to the decision.
Politico’s article does not disclose Ecosia’s request volumes, failure rates, latency measurements or service-level commitments. That limits what outsiders can conclude about the infrastructure dispute, while leaving the commercial significance of the complaint intact: a customer has publicly said the combination was unacceptable.
Mistral Has Offered Another Test, Not a Confirmed Resolution
Mistral did not respond to Politico’s request for comment for the article, according to the publication. The report does include remarks from co-founder and chief scientist Guillaume Lample at a press conference encouraging Ecosia to try a newer model.
“I would encourage them to test our new model as soon as possible. We can even give them early access today if they want.”
Lample also pointed to Mistral’s ability to provide model weights and customize models with customers. That gives Ecosia a possible route to reassess the relationship by evaluating newer technology and considering a different deployment or customization arrangement.
There is no confirmation that Ecosia accepted the offer, that another evaluation occurred or that the reported server problems were resolved. The report supplies no results demonstrating how the newer model performs on Ecosia’s tasks.
In a fast-moving AI industry, a customer may reasonably reject the service it has been receiving while a supplier argues that its next release will address the shortcomings. Lample’s offer belongs in the account, but the prospect of improvement does not establish that the existing complaint has been answered.
Ecosia’s Help Page Describes Two Different AI Features
Ecosia’s documentation names specific models for two distinct features.
The help page says AI Overviews use Mistral Small 3.2. These overviews summarize information from search results and provide source links above the results page. It also says the feature is powered by Staan through EUSP.
For the conversational experience, AI Search uses Mistral Small 4. This feature delivers full-text answers and lets users explore a topic through follow-up questions. The documentation calls it both “AI Search” and “AI Chat.”
The page does not describe identical provider arrangements for these two product surfaces, so a single supplier decision cannot safely be assumed to change both at the same time.
The help page could be outdated, the migration could be incomplete, or the reported switch could cover only part of Ecosia’s AI offering. None of those explanations is established by the available evidence.
For now, Ecosia’s CEO says the company is dropping Mistral, while its public documentation continues to identify Mistral models. It would be premature to say that all Ecosia AI responses now come from Chinese models, or that Mistral has already disappeared from every part of the product.
Users cannot settle the question from their own experience alone. A changed answer style or a faster response would not, by itself, identify the model responsible. A feature-by-feature provider disclosure would be more useful than guesses based on outputs.
The Replacement Platform Is Named, but the Models Are Not
According to Politico’s account of the replacement arrangement, Ecosia is partnering with Melious, a platform that hosts open-source and open-weight models. The report lists Qwen, GLM and Kimi among the Chinese model families available through that platform.
A platform’s catalog does not establish a customer’s production configuration. The report neither confirms that Ecosia selected any of those families nor identifies the exact replacement models or explains how requests are routed among them.
Kroll claimed that the new arrangement improves the economics:
“We’ve roughly cut our costs in half while improving quality and performance.”
That cost reduction has not been independently verified. The report gives no original spending baseline, traffic levels, contract terms or controlled comparison. It also leaves the quality and performance measurements undefined.
Costs depend on more than the model name. A fair comparison would need to account for workload, response length, hosting arrangements and the amount of computation used per answer. A lower bill alone would not reveal which part of the system became cheaper.
The reported move still illustrates a meaningful form of competition. Access to model weights can give customers more options over who operates a model, subject to licensing and technical requirements. A company dissatisfied with one hosted service may be able to change both its model supplier and its infrastructure provider.
That flexibility is part of the appeal Kroll describes, and it complicates any reading of the switch as a simple head-to-head contest between two model developers.
European Independence and Model Bias Remain Separate Questions
Kroll’s argument extends beyond answer quality. He told Politico that Europe can build on increasingly capable open models rather than spending billions to train frontier systems itself. He also questioned whether Mistral’s international investors align with Ecosia’s idea of technological sovereignty.
Sources
- Politicopolitico.eu
- AI Search help pagesupport.ecosia.org
