Ecosia Changed Its AI Twice in Five Months. Could You?

In May, Ecosia moved its AI features from OpenAI to Mistral. Five months later, it moved again.

Two big model switches in one year. For a search engine. And by the CEO's account, the second one made things cheaper and better at the same time.

Most teams I know would need a quarter just to agree on the migration plan.

The short version

Ecosia, the Berlin search engine that plants trees, dropped Mistral in early October. Its AI chat and search summaries now run on open-weight models like Qwen, GLM and Kimi, served by Melious AI, a German startup that runs open models in European data centers.

CEO Christian Kroll said they roughly halved their AI costs while improving quality. His reasons for leaving were not gentle. He called Mistral's quality "a year behind", complained about overloaded servers, and questioned how sovereign Mistral really is with investors like Samsung and Nvidia. He even mentioned the nuclear share in France's energy mix. Mistral's CEO Arthur Mensch answered on X: "Sorry if using nuclear energy is a real blocker."

Europe's tech drama has a dry sense of humor.

Then this Thursday, October 29, Mistral released the weights of Large 4, its roughly one trillion parameter model. Internal use is free. If you want to sell it as an API to others, you need a separate commercial license.

So the model Ecosia left is now something Ecosia could, in theory, run itself. That's how fast this moves.

The model is just a dependency now

Here is what I keep thinking about. Nobody at Ecosia wrote a heroic rewrite. They could switch twice because, somewhere in their system, the model was a replaceable part.

As a backend engineer, this feels very familiar. In Spring Boot we don't marry a database. We talk to it through a repository, and on a good day you could change the driver without the business logic noticing. Payment providers, email services, cloud storage: we learned the hard way to keep them behind an interface.

With AI, a lot of teams forgot that lesson. The prompt is tuned for one model's quirks. The response parsing assumes one vendor's format. The SDK is imported in forty files. And there are no tests that tell you if a new model is better or worse for your use case. Only vibes.

Loyalty to a model is not a strategy. The ability to leave it is.

What I expect next

More switching, not less. Open-weight models keep getting closer to the closed ones, European hosting is getting easier, and prices keep moving. The team that can change providers in two weeks gets to negotiate. The team that needs six months just pays the invoice.

I also expect licenses to matter more. "Open weights" does not mean "do whatever you want", as the Large 4 license shows. Someone on your team has to actually read it.

What we can do this week

You don't need a big platform project for this. A few small habits go a long way:

1. Put one interface between your code and the model. Your services should call something like SummaryService, not a vendor SDK directly.
2. Build a small eval set. Fifty real inputs with expected good answers, from your own product. That's your test suite for any new model.
3. Track cost per feature, not per month. "Summaries cost us X per thousand requests" is a number you can compare across providers.
4. Keep prompts in config, versioned per model. When you switch, you will need to tune them, and you want that diff to be visible.
5. If you lead a team, ask one question in your next planning: "If our model provider doubled the price tomorrow, how long until we could move?" The silence after that question is useful information.

Ecosia did not win because they picked the perfect model. They won because they could stop picking.

Sources: Forbes: Europe's AI Race Is Moving Beyond Who Builds The Biggest Models, Dealroom: Ecosia drops Mistral for open-weight models, Developers Digest: Mistral Large 4 weights and license, DEV Community: Mistral Large 4 weights are out