Ecosia Drops Mistral for China's Open Models

Ecosia, the Berlin-based search engine that puts its profits toward tree planting, has dropped Mistral as its AI supplier after about five months and moved to a set of open-weight models from Chinese labs. Chief executive Christian Kroll told POLITICO the company was "disappointed with the quality of Mistral" and described the French lab as "a year behind" its competitors, citing weak results and recurring server overloads.

In place of Mistral, Ecosia now draws on open-weight models including Alibaba's Qwen, Zhipu AI's GLM, and Moonshot AI's Kimi. Rather than contract with any single lab, it routes requests through Melious, a German platform that hosts the models and serves them over an API. Melious says data never leaves Europe, which lets Ecosia use Chinese model weights without sending European user queries to servers in China.

Kroll said the switch roughly halved Ecosia's costs while improving quality and speed. That figure comes from Ecosia and has not been independently checked. The company had only adopted Mistral in May, after moving away from OpenAI, so this is its second supplier change inside a year.

Open weights, awkward politics

The decision lands on a sore spot. Ecosia is used by a number of European public authorities, and dropping Europe's flagship AI lab in favour of Chinese models cuts against the continent's push for technological sovereignty. Mistral has been held up as proof that Europe can field a frontier lab of its own. A prominent European customer walking away, and saying the product was not good enough, is not the endorsement that argument wanted.

The timing stung too. Mistral had just unveiled Large 4, its most powerful model yet, when Ecosia's decision surfaced. Co-founder Guillaume Lample publicly invited Ecosia to test the new release, which suggests the comparison had been made against older models.

What the move really shows is how good the open-weight Chinese models have become, and how cheap. A year ago, a European company swapping a domestic commercial model for Qwen or GLM would have meant trading quality for price. Kroll's claim is that he got both at once: lower cost and better output. Open weights from labs such as StepFun have been closing the gap with the frontier through the year, and they can be run from European soil, which sidesteps the data-residency worry that usually follows anything labelled "made in China."

For users, the change is meant to be invisible; the search results and AI features carry on as before. For the wider market, it is a small but telling data point. When a values-driven European firm with every political incentive to buy local instead buys open weights from Hangzhou and Beijing, the reason tends to be simple. The models were better for the money.