- Ecosia replaced OpenAI with Mistral in May and, its CEO told Politico, is now moving to open-weight models, including Chinese ones, saying costs fell by about half.
- The claims are the CEO's own, and the sources do not say how hard the switch was or whether it is finished. Bias in Chinese models is a disputed risk.
- Ask your team how fast you could change AI supplier, what your contract says about capacity, and who tests the model for bias.
One company's switch raises a question about portability
In May, the Berlin-based search engine Ecosia replaced OpenAI with France's Mistral as its AI provider. Now it is dropping Mistral for open-source and open-weight models, CEO Christian Kroll told Politico. Politico gives no date for when the move will be complete.
The lesson here is a question, not a verdict. If the model behind a product can be swapped, it is a part rather than a partner. Ecosia's story does not show that swapping is easy. The source does not say how hard the move was, or what Ecosia's contracts and code allow. That is the question to put to your own team.
Ecosia's reach makes the move worth noting. Politico reports that institutions from Germany's federal environment ministry to the U.K.'s National Health Service use it. This is one company's decision, not a market trend.
What Ecosia says went wrong
Kroll said he is "disappointed with the quality of Mistral". He said its models are now "a year behind" the competition. These are his views, and Politico did not report independent tests.
He also described recurring technical problems, including overloaded servers. "We were simply too large a customer for Mistral," he said. For a buyer, that is a capacity problem as much as a quality one.
Kroll raised a third concern: sovereignty. He said Mistral's reliance on international investors is, in his view, "not truly sovereign". Politico notes that Mistral has raised billions of euros abroad, including from Samsung Electronics and Nvidia.
Mistral did not respond to Politico's request for comment. Its chief scientist, Guillaume Lample, told Politico at a Monday press conference that Ecosia should test Mistral's "new model" and offered early access. The next day, Mistral released a new model, Large 4. The article does not say whether Ecosia has tested any Mistral model since.
How open-weight models change the deal
A model's weights are the learned numbers that make it work. With an open-weight model, those numbers are published. Anyone can run the model on their own servers or on a host.
Ecosia is partnering with Melious, a platform that hosts such models. They include the Chinese models Qwen, GLM and Kimi. Kroll said, "We've roughly cut our costs in half while improving quality and performance." That figure is his own. The article does not report any outside check.
Mistral argues it also offers a route to control. Lample said it can give customers the weights of its models, large and small, and customise them together. So the real choice is not only which lab to pick. It is who runs the model, who holds the weights and who can change them.
The bias question is a testing question
Chinese models carry a disputed risk. Politico points to a recent NewsGuard investigation of the five leading Chinese-backed models. In it, the models repeatedly let false pro-China claims go unchallenged. Western research groups have also doubted how well the models work, citing their censorship of politically sensitive topics.
Kroll says such distortions are not a dealbreaker. In his view, technical measures can address them, unlike weak model quality. Experts disagree in part. Rasmus Rothe of the German AI Association says refusals and evasive answers can largely be removed by targeted retraining.
Rothe says the harder problem is quiet bias. It is built into the data a model learns from and the assumptions it carries. A few hundred test questions, he said, will only "scratch the surface".
Questions to put to your team
Start with portability. If your main AI provider failed tomorrow, how long would a switch take? We do not know how hard Ecosia's swap was. Ask whether your own contracts and code would allow one.
Ask about capacity. Does your contract say what happens when you become a large customer? Overloaded servers are a business risk, not only a technical one.
Ask who tests for bias, and how. Rothe's point is that quick fixes may reach only the surface. Any model used in customer-facing answers needs its own test questions, run before launch and again after each model change.
Finally, define sovereignty in writing. Ecosia questions it for Mistral's investors. Chinese models bring their own concerns, including U.S. worries about national security and Western research organisations' doubts about their censorship. Decide which risks your organisation can accept, then check each supplier against that list.
A model you can swap may be cheaper to leave and easier to question. It still has to be tested before you trust it.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Politico.
