Ecosia drops Mistral for open-weight models including Qwen, GLM and Kimi: Europe's 'sovereign AI' shifts to who runs the model

Why it matters

【Key points】

Berlin search engine Ecosia is dropping France's Mistral for open-weight models, including China's Qwen, GLM and Kimi. CEO Christian Kroll told POLITICO on Oct 6 he was "disappointed with the quality of Mistral," calling the models "a year behind." Ecosia only switched from OpenAI to Mistral in May, then hit server overloads: "We were simply too large a customer for Mistral." He also cited France's nuclear-heavy power mix and foreign investors in Mistral as "not truly sovereign."

The new setup runs through Melious: open-weight models only, inference routed across EU providers. Kroll says costs roughly halved while quality improved—his claim, with no published eval. Ecosia's users include Germany's environment ministry and the UK's NHS. Mistral co-founder Guillaume Lample invited Ecosia to try Large 4 early and said Mistral can hand over weights and customize with customers.

【Verifiable】

POLITICO (Oct 6): https://www.politico.eu/article/germany-ecosia-search-engine-mistral-china-open-source-ai/

Melious docs: https://melious.ai/docs

【Our read】

1. This is less "Chinese models beat Europe" than a shift in what sovereignty means: from who trains the model to who runs it and holds the weights. A Europe-independence brand can run Chinese open weights in EU data centers and call that more sovereign.

2. Capacity hurt Mistral before quality did. When a closed API can't serve a large customer, buyers move to open weights they can host elsewhere. Lample's reply (hand over weights, customize) concedes the same logic.

3. Two supplier switches in five months show the model layer becoming a swappable part; leverage moves to routing layers like Melious. "Half the cost" is so far the CEO's word.

4. Risk remains. Rasmus Rothe of the German AI Association notes overt censorship can be retrained away, but subtle bias in training data is harder. A search engine used by public institutions should publish how it tests for it.