Ecosia Drops Mistral for Open Models Hosted in the EU

What Ecosia Said, And What Mistral Answered

On 6 October 2026 Politico reported that Ecosia, a Berlin search engine used by institutions from Germany's federal environment ministry to Britain's NHS, is dropping Mistral for open models, including Chinese ones. Ecosia had replaced OpenAI with Mistral only in May.

CEO Christian Kroll gave three reasons. He is disappointed with the quality of Mistral's models, which he calls a year behind the competition. Ecosia ran into recurring technical problems, including overloaded servers, because it was simply too large a customer for Mistral. And in his view reliance on international investors is not truly sovereign. Mistral has raised billions of euros abroad, including from Samsung Electronics and Nvidia.

Mistral did not respond to Politico's request for comment, but its chief scientist Guillaume Lample told the outlet he would encourage Ecosia to test the new Large 4 as soon as possible, and offered early access, the weights and joint customisation.

Where The Models Stand On One Index

On the Artificial Analysis Intelligence Index, as reported by Trending Topics, Mistral Large 4 scores 38.4, the strongest open model outside China but only eighth among open models, behind seven Chinese ones.

Per task, Large 4 costs about $1.13, against about $2 for GLM-5.3 and Kimi K3 and $0.27 for DeepSeek V4.1 Flash, which scores slightly higher. Two caveats apply. Large 4 is a preview still in training, with weights due at the end of October, and Mistral says it was trained entirely in Europe and can run on customers' own servers. And Ecosia's help page lists Mistral Small 3.2 and Small 4 for its AI features, not Large, so the model Kroll judged was not the one Mistral launched that week.

What Melious Is Selling

Ecosia now gets its models through Melious, a German platform that, by its own description, offers more than 60 open-weight models through one API compatible with OpenAI and Anthropic code, so a switch needs only a new base URL and key. It says its 11 providers sit in eight countries, Luxembourg, France, Italy, Ireland, Sweden, Germany, the Netherlands and Finland, that automatic failover is built in, that your data never leaves Europe and that the setup was independently checked by staysin.eu. These are homepage claims we could not verify.

The lineup includes Z.ai's GLM 5.3, Moonshot AI's Kimi K3 and DeepSeek V4.1 Flash. Trending Topics notes that Z.ai has been on the US entity list since 2025, and that Europeans who want top open models almost inevitably end up with Chinese developers, even when the servers are European.

Kroll says Ecosia has roughly cut its costs in half while improving quality and performance. No figures were published, and both Politico and TechNode relay the claim from him.

A Sovereignty Checklist For Your Next AI Contract

The case moves the question from who built the model to where it runs and how fast you can change it. A European data centre can host a Chinese model, which is exactly what Ecosia chose.

Check four things. Where inference runs and which law applies to it. Who operates the servers, since eleven providers behind one API means you need the list in the contract. How long a model swap takes, which you should test rather than read: Ecosia went from OpenAI to Mistral in May and off Mistral in October. And where the model comes from, because Politico cites a NewsGuard finding that five leading Chinese-backed models repeatedly failed to correct pro-China falsehoods, and a US entity-list vendor can matter if you sell into the US.

Treat any cost claim as a hypothesis. Cost per task differs by a factor of four between Large 4 and DeepSeek V4.1 Flash on the index, so run your own task-level comparison before moving.

Read next: 70% of Europe's $25B Went to 83 Late-Stage Companies  |  A Free AI Model Just Beat What You Pay For Research