German search engine Ecosia is replacing Mistral with open-weight AI models, including Chinese-developed Qwen, GLM and Kimi, through a partnership with German AI platform Melious. Founder and CEO Christian Kroll said the switch had roughly halved costs while improving quality and performance.
Can’t argue with half the cost. My employer (a large corpo) has also deployed GLM internally and the usage is unlimited, while Big Tech’s LLMs are heavily limited now.
I don’t thing this trend would reverse. There’s no material reason to. Countries who want their own models would have to put model development under public institutions or provide public funding for private ones to do it. As the American AI bubble deflates, the pressure for using free open models would only increase.
Will be interesting to ask it a couple questions about Taiwan or Tiananmen square… In principle, the censorship can be edited out of open-weight models, but I’m not sure whether Melious does that.
Well, don’t know if they do, but that’s the general advantage of open weight. You control infrastructure, you know exactly what happens to your data, prompts, context windows. If you so wish, you can finetune the model, even just LoRA so the cost is pretty low too. Totally feasible for local companies that own a few A100s/H200s/MI350s.
Mistral, anthropic, openai: no chance, pay up buttercup, and if your data is interesting to us we’ll steal it and publish your work as our own
Closed models are a huge risk to use for anything of value
So Qwen is trained on Chinese propaganda. What is GLM and what is Kimi?
Both Chinese. But let’s not forget that American models are trained on American and Israeli propaganda too.
This is sad news,but I’ve used Mistral until fairly recently, until they changed the way tokens are being counted or some change around providing “more transparency”, resulting in barely getting any usage out of my 18€ pro subscription.
Yeah, I tried to support more of EU economy,but I was throwing my money out the windows, compared to their competitors and their models and offers.
I wish they could see they are not competitive at all.
Edit: I don’t think going full Chinese LLM is the answer.
Mistral was already way behind the Chinese models let alone the American ones. They were pretty good in… 2025.
improving quality and performance.
By which measures? Spreading Chinese propaganda can never improve quality nor performance.
Ecosia adopted Mistral in May after moving away from OpenAI, but Kroll said it encountered recurring technical problems, including overloaded servers, and was disappointed with model quality.
Apparently by the measures of availability and quality of output, since that is what they said they were disappointed by. But reading is kinda hard isn’t it?
How can propaganda slop lead to better quality of output? This is impossible.
They were on ClosedAI before, so that’s basically US/Israeli propaganda slop. They only used Mistral for a few months and it turned out Mistral just doesn’t work much of the time. China is the only remaining player.
Because close to all prompts users put into ecosias search do not ask for information that China has a bias towards?
China has bias towards a lot more than immediate China-related topics. You can’t trust such a regime.
