Ecosia, the Berlin-based search engine that funnels ad revenue into tree-planting projects, has dropped Mistral as its AI model supplier and moved to a trio of open-weight models instead. The switch, reported by TechNode and confirmed by multiple outlets through October 21, 2026, hands Ecosia’s AI search features to Alibaba’s Qwen, Z.ai’s GLM, and Moonshot AI’s Kimi, all served through a partnership with German AI platform Melious. Founder and CEO Christian Kroll says the move roughly halved his company’s AI costs while improving output quality, a claim that has since rippled through the European tech press as a case study in the continent’s growing appetite for open-weight alternatives to proprietary providers.
The story matters beyond one mid-sized search engine’s vendor list. Ecosia is used by Germany’s federal environment ministry and the UK’s National Health Service, according to a DeAI.org report on the switch, giving its procurement decisions a visibility that a typical consumer app would not carry. It also lands in the middle of a broader argument that has consumed European tech policy circles all year: whether the continent’s AI stack should run on homegrown or allied models, Chinese open-weight releases, or the usual slate of US and French proprietary systems. Ecosia just answered that question with its wallet, and the answer was not Mistral.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Ecosia Ends Its Five-Month Mistral Experiment
Ecosia’s relationship with Mistral was short. Per reporting from TechFlier, the search engine adopted Mistral’s models around May 2026, after moving away from OpenAI. That arrangement lasted roughly five months before Kroll began shopping for a replacement. Five months is a short runway for an infrastructure decision that touches a company’s core product, and it signals how quickly dissatisfaction built up once Mistral’s models went into production at Ecosia’s scale.
Kroll was blunt about why the partnership ended. He said Mistral was “a year behind” rivals and that Ecosia had become “simply too large a customer” for Mistral’s servers to handle reliably, according to the quotes reproduced in DeAI’s reporting on the switch. Those are not the kind of lines a company usually puts on the record about a vendor it hopes to work with again, and they suggest the breakup involved both a performance complaint and a capacity one. Ecosia, in other words, says it outgrew Mistral’s infrastructure at the same time it lost confidence in the quality of Mistral’s output.
Why Kroll Says Mistral Fell Behind
The quality complaint is the more pointed of the two. Kroll told reporters he was “disappointed with the quality of Mistral,” a direct criticism of France’s flagship AI lab from a customer using its models in a live consumer product. Mistral has spent 2026 positioning itself as Europe’s answer to US and Chinese labs, with backing from the French government and EU institutions eager to have a domestic champion in the foundation-model race. A public rebuke from a mid-sized but visible European customer cuts against that narrative, regardless of how representative Kroll’s experience turns out to be across Mistral’s broader customer base.
It is worth separating the two complaints Kroll raised. “A year behind” is a claim about model capability relative to competitors, the kind of subjective judgment that depends heavily on which benchmarks and use cases you weight most. “Simply too large a customer” is a claim about operational capacity, which is a more concrete and checkable kind of failure. Search engines generate huge, bursty request volumes, and an AI vendor that cannot scale cleanly with a client’s traffic creates outages and latency spikes that are hard to paper over. If that capacity complaint is accurate, it says as much about Mistral’s infrastructure investment as it does about its model quality.
Inside the Melious Deal: EU-Hosted Open-Weight Models
Rather than signing with another single proprietary vendor, Ecosia routed its new AI workloads through Melious, a German AI platform that hosts open-weight models on servers spread across eight EU countries, according to the DeAI Daily Brief covering the announcement. That detail is the real headline for anyone tracking European data sovereignty debates: Ecosia did not simply swap one closed model for another, it moved to a hosting layer that keeps inference inside the EU while running models whose weights were developed in China.
That combination, open-weight models with Chinese origins served on EU-based infrastructure, is becoming a recognizable pattern. It lets a European company sidestep both the regulatory exposure of sending data to US-run clouds and the commercial dependency of relying on a single national lab’s capacity. Melious effectively functions as a sovereignty-friendly middleman: European hosting, but model choice unbundled from any single country’s foundation-model output. Whether that pattern becomes the default playbook for other European companies weighing similar decisions is one of the more interesting open questions the Ecosia story raises.
Qwen, GLM and Kimi: Who Made These Models and Why They Matter
The three models Ecosia picked all come from different Chinese labs, each with its own open-weight release strategy that has made them attractive to developers outside China looking for a cheaper or more controllable alternative to closed US models.
Qwen (Alibaba)
Qwen is Alibaba’s open-weight model family, one of the most widely deployed outside China thanks to a steady release cadence and permissive licensing terms that make it straightforward for third parties like Melious to self-host.
GLM (Z.ai) and Kimi (Moonshot AI)
GLM comes from Z.ai, formerly known as Zhipu AI, a lab that has built a reputation for competitive reasoning and coding performance in open-weight form, a trajectory Tech-Insider has previously covered in the context of third-party safety testing on GLM’s recent releases. Kimi is Moonshot AI’s model line, which has carved out a niche with long-context handling and aggressive open-weight releases of its own. None of the three were built with Ecosia’s search product specifically in mind, which is itself notable: Ecosia is adapting general-purpose open-weight models to a search-overview use case rather than waiting for a purpose-built search model.
| Model | Developer | Country of origin | License approach | Role at Ecosia |
|---|---|---|---|---|
| Qwen | Alibaba | China | Open-weight | AI search overviews |
| GLM | Z.ai (formerly Zhipu AI) | China | Open-weight | AI search overviews |
| Kimi | Moonshot AI | China | Open-weight | AI search overviews |
| Mistral (previous supplier) | Mistral AI | France | Mixed open/proprietary | Dropped after ~5 months |
| OpenAI (earlier supplier) | OpenAI | United States | Proprietary | Dropped before Mistral adoption |
From OpenAI to Mistral to Open-Weight: Ecosia’s Supplier Timeline
Laid out chronologically, Ecosia’s AI supplier history looks less like a single decision and more like a company still searching for a stable footing. It moved away from OpenAI at some point before May 2026, adopted Mistral in May 2026, and by October 2026 had moved again, this time to the Qwen, GLM, and Kimi combination hosted via Melious. Three supplier changes in under two years is a fast churn rate for any company’s core AI infrastructure, and it suggests Ecosia has been willing to treat model selection as a live, ongoing negotiation rather than a settled architectural choice.
| Period | Supplier | Approx. duration | Reported reason for change |
|---|---|---|---|
| Before May 2026 | OpenAI | Unspecified | Moved away prior to Mistral adoption |
| May 2026 – Oct 2026 | Mistral AI | ~5 months | Quality concerns; capacity strain at Ecosia’s scale |
| Oct 2026 onward | Melious (Qwen, GLM, Kimi) | Ongoing | Lower cost, EU-hosted open-weight alternative |
The Cost Math Behind “Roughly Cut in Half”
Kroll’s central claim, that the switch “roughly cut our costs in half while improving quality,” is the figure that made this story travel. It is worth noting, as DeAI’s own coverage of the interview does, that this is a self-reported number from the company making the switch rather than an independently audited cost breakdown. That does not make it false. Open-weight models hosted on shared European infrastructure are generally cheaper per token than frontier proprietary APIs, and Ecosia would have direct visibility into its own invoices either way. But a 50 percent savings figure from the party with every incentive to publicize a successful vendor switch deserves the same healthy skepticism any company’s self-reported efficiency claim would get in a product earnings call.
What makes the claim plausible on its face is the structural difference between how proprietary frontier labs price API access and how open-weight hosting platforms like Melious typically price theirs. A single-vendor proprietary contract at Ecosia’s traffic volume carries premium pricing tied to that vendor’s own infrastructure and margin requirements. A multi-model open-weight hosting arrangement lets the hosting platform shop for the most cost-efficient compute and pass savings through, and it lets Ecosia switch models without renegotiating a master contract every time. Whether that translates into exactly half the cost is unverifiable from the outside, but the direction of the claim is consistent with how the broader open-weight hosting market has been pricing itself against closed-model incumbents this year.
Why Chinese-Developed Open-Weight Models Are Gaining Ground in Europe
Ecosia’s move did not happen in a vacuum. European companies have spent 2026 quietly testing Chinese open-weight models as a hedge against both US cloud dependency and the slower release cadence of domestic labs like Mistral. The appeal is straightforward: Qwen, GLM, and Kimi are free to download, cheap to host, and improving fast, while a European company retains full control over where the inference actually runs. That last point is what Melious is selling, infrastructure that keeps the data path inside the EU even when the model weights themselves were trained by a Chinese lab.
It is also a trend that cuts against a simpler, more comfortable narrative in Brussels, that Europe’s AI future runs through European-built models. Ecosia’s decision is a visible data point suggesting that, when cost and reliability are on the line, at least some European companies are prioritizing open-weight access over country of origin. That tension between AI sovereignty as a policy goal and AI sovereignty as a purchasing decision is one Tech-Insider has tracked in coverage of how think tanks are framing the stakes of Europe’s position in the broader AI race, and the Ecosia case gives that debate a concrete, named example rather than an abstract policy argument.
Competitive Landscape: Mistral vs the Open-Weight Alternatives
Mistral has not issued a public response to Kroll’s comments, based on available reporting, which leaves this as a one-sided account for now. But the competitive pressure on Mistral is real regardless of how one customer’s complaint lands. Mistral spent much of 2026 trying to differentiate itself on both performance and “sovereign AI” positioning, the idea that European companies should prefer a European-built model on principle. Losing a visible, climate-branded European customer to Chinese open-weight alternatives undercuts both halves of that pitch at once: it is not winning purely on capability, and it is not winning purely on home-continent loyalty either.
The broader open-weight field Ecosia drew from has also gotten more crowded and more capable in recent months. DeepSeek pushed out an MIT-licensed update to its V4.1-Flash model in October, and the same week as the Ecosia story broke, DeepSeek also open-sourced the production inference framework it uses to serve that model at scale, a move aimed squarely at lowering the cost of running open-weight models for exactly the kind of mid-sized company Ecosia represents. Reflection AI’s open-weight Beam model, which Tech-Insider covered at launch, is another entrant in the same race to give companies a credible non-proprietary option. Mistral, by contrast, has built its reputation partly on keeping a foot in both the open and closed camps, a strategy Tech-Insider examined when covering Mistral’s Timestral forecasting model earlier this year. Ecosia’s defection suggests that straddling both camps is getting harder to defend when fully open-weight competitors are improving this quickly.
Historical Context: Europe’s Long Search for AI Sovereignty
Europe’s push for AI independence did not start with Ecosia and will not end with it. Mistral itself was founded in 2023 largely as an answer to the dominance of OpenAI and Google, with French government support and EU-friendly messaging baked into its identity from the start. The hope among European policymakers was that a homegrown champion would reduce the continent’s reliance on US infrastructure while keeping pace on capability. Three years later, one of Mistral’s higher-profile European customers has concluded that pace was not fast enough, and has turned instead to a hosting-layer solution that sources its models from China while keeping the servers on European soil.
That is a meaningfully different outcome than the one European AI policy was built around. The original sovereignty argument assumed the choice was between European models and American ones. Ecosia’s move shows a third path emerging, European-hosted infrastructure running internationally sourced open-weight models, that policymakers in Brussels and Berlin had not fully priced into their planning. It is a reminder that open-weight licensing has scrambled the usual national-champion logic that shaped Europe’s AI strategy earlier in the decade.
Market Impact: What This Means for Mistral, Melious and the Open-Weight Race
For Mistral, the immediate financial impact of losing Ecosia’s business is almost certainly modest; Ecosia’s traffic, while meaningful by search-engine standards, is unlikely to be material to Mistral’s overall revenue. The reputational impact is the bigger risk. Mistral has spent the year courting exactly the kind of mid-sized European company Ecosia represents, companies that want European infrastructure but need enterprise-grade reliability and competitive pricing. A public, on-the-record complaint from inside that target customer base is the kind of story that shows up in other companies’ procurement conversations long after the headlines fade.
For Melious, the opposite is true. Landing a known, publicly branded customer like Ecosia is a meaningful validation for a hosting platform still building its reputation outside Germany. If Melious can point to Ecosia’s cost claims as a reference case, even a self-reported one, it gives the platform a marketing asset other European companies evaluating open-weight hosting will take seriously. For the broader open-weight ecosystem, the signal is that European companies are now comfortable enough with Chinese-developed open-weight models, when paired with EU-based hosting, to make them production infrastructure rather than just experimentation targets.
The Data Residency Angle: Why “EU-Hosted” Matters
The detail that Melious runs its infrastructure across eight EU countries is not incidental to this story, it is arguably the load-bearing fact. European companies serving government and healthcare clients, as Ecosia does with Germany’s environment ministry and the UK’s NHS, face data residency and procurement requirements that a US-hosted API cannot always satisfy cleanly. By routing Chinese-developed model weights through EU-based servers, Ecosia and Melious appear to be threading a specific regulatory needle: comply with EU data residency expectations on the infrastructure side, while sourcing model capability from wherever it is cheapest and most competitive on the model side.
That split matters because most of the public debate over Chinese AI models in Europe and the US has focused on where the model was trained and by whom, rather than where the inference actually runs. Ecosia’s setup suggests those two questions can be decoupled in practice, which could become a template other security-conscious European buyers lean on if they want access to fast-improving open-weight models without routing sensitive queries through non-EU infrastructure.
Risks and Open Questions: The FAQ Still Lists Mistral
One detail undercuts the tidiness of the announcement: according to a report from NeoTeo, Ecosia’s own help pages and FAQ were still naming Mistral as the model behind its AI Overviews feature even after the switch was publicly announced. That is a minor operational detail, almost certainly a documentation lag rather than evidence the switch has not actually happened, but it is a useful reminder that announced infrastructure changes and fully rolled-out infrastructure changes are not always the same thing on the day of the press cycle. Companies routinely announce a switch before every corner of their public-facing product has caught up to it.
It also raises a fair question about how completely Ecosia has migrated. Search products often run multiple models in parallel during a transition, for testing, fallback, or gradual rollout reasons, and outside reporting has not confirmed whether Mistral has been fully removed from Ecosia’s stack or is still running in some reduced capacity behind the scenes. Readers should treat “Ecosia dropped Mistral” as an accurate description of the company’s stated direction, not necessarily as proof of a fully completed migration on the exact date of the announcement.
What This Signals for Other Mid-Sized AI Buyers
Ecosia is a useful case study precisely because it is not a hyperscaler or a frontier lab, it is a mid-sized company making the same kind of build-versus-buy decision that thousands of other European businesses face when they add AI features to an existing product. Its public willingness to name names, criticize a vendor on the record, and publish a specific cost-savings figure is unusually candid for this kind of procurement story, which is exactly why it traveled across multiple outlets within days.
Other companies watching this unfold are likely drawing a simple lesson: open-weight models hosted on EU infrastructure are now a credible production option, not just a cheaper fallback for non-critical workloads. Combined with the pace of recent open-weight releases from labs like DeepSeek and Reflection AI, the practical gap between “use the open-weight option to save money” and “use the open-weight option because it is simply competitive” appears to be narrowing faster than Europe’s slower-moving domestic labs can match.
Predictions: Where This Trend Goes Next
- Expect more mid-sized European companies to publicly disclose similar open-weight migrations over the coming months, especially ones already running cost-sensitive, high-traffic consumer products like search or recommendation engines.
- Mistral will likely respond with pricing or performance commitments aimed at enterprise and public-sector customers specifically, the segment Ecosia’s complaint was most damaging to.
- Melious and similar EU-hosted open-weight platforms are likely to use the Ecosia case as a reference customer in sales conversations with other European companies weighing the same vendor question.
- The “EU-hosted, internationally sourced models” pattern is likely to draw more scrutiny from European policymakers, who will need to decide whether data residency alone satisfies their sovereignty goals or whether model origin still matters independently.
- Expect continued ambiguity over self-reported cost-savings figures like Ecosia’s, since few companies making vendor switches have an incentive to commission independent audits of their own claims.
What Tech-Insider Readers Should Watch Next
The most useful signal going forward is whether Ecosia’s FAQ and public documentation catch up to the announcement in the coming weeks, which would confirm the migration is complete rather than partial. The second signal worth tracking is whether Mistral responds publicly at all, since silence from a company usually eager to promote its European credentials would itself be notable. And the third is whether other companies using Melious or similar EU-hosted open-weight platforms come forward with their own cost and quality comparisons, which would turn Ecosia’s single data point into an actual trend line rather than an isolated anecdote.
Frequently Asked Questions
What did Ecosia actually announce?
Ecosia said it is replacing Mistral AI with open-weight models, Alibaba’s Qwen, Z.ai’s GLM, and Moonshot AI’s Kimi, served through a partnership with the German AI hosting platform Melious.
Why did Ecosia drop Mistral?
CEO Christian Kroll said he was disappointed with Mistral’s model quality, calling it “a year behind” competitors, and said Ecosia had become “too large a customer” for Mistral’s infrastructure to handle well.
How long did Ecosia use Mistral?
Roughly five months. Reporting indicates Ecosia adopted Mistral’s models around May 2026 and began moving away from them by October 2026.
How much money did Ecosia save by switching?
Kroll said the switch “roughly cut our costs in half” while also improving output quality. That figure is self-reported by Ecosia and has not been independently verified.
What is Melious?
Melious is a German AI platform that hosts open-weight models, including Qwen, GLM, and Kimi, on servers located across eight EU countries, giving European companies a way to use internationally developed models while keeping inference on EU-based infrastructure.
Are Qwen, GLM, and Kimi Chinese models?
Yes. Qwen is developed by Alibaba, GLM by Z.ai (formerly Zhipu AI), and Kimi by Moonshot AI, all Chinese AI labs that release their models under open-weight licenses.
Has Mistral responded to the criticism?
Based on available reporting, Mistral has not issued a public statement responding to Kroll’s comments or to Ecosia’s decision to switch suppliers.
Is Ecosia fully off Mistral now?
Not entirely confirmed. A report from NeoTeo noted that Ecosia’s own FAQ pages were still naming Mistral for its AI Overviews feature after the switch was announced, suggesting the migration may still be rolling out across the product.
