Germany's NEW AI Model is INSANE (FREE!)

Julian Goldie (0:00)

Germany just dropped a brand new AI model called Calibri, and it's completely free. You can download it right now and run it on your own computer locally. No account, no sign up. It's yours. And here's the exciting part. On the benchmarks, it's beating big names like Quen, Nemetron and Mistral. In this video, I'm going to show you what this model can do, where to get it, and the one thing you absolutely have to know before you download it. And I'll show you that later. So see me to the end, because by the end of this video, you could have your own private AI running locally for free on your own machine, working around the clock, 24-7, completely free. Let's get into it.

So today we have a brand new model from Germany, and it's from Aleph Alpha. That's the company that made it. The model itself is called Colibri, and it is described as a small bird with fast wings, 78 billion parameters, 3.46 active, up to 1 million token context window, and it was built in Europe of all places. Now, aside from mistrial, there's not really been many serious models coming out of Europe. So this is pretty exciting to see. And this is open weight. So you can get it. You can run it on your own hardware. It's under the Apache 2 license as well. And you can get it directly on Hugging Face. I'll run through the parameters and also the technical report in a minute. So you can see the technical report over here. And then also we have the details on here. So yeah, it came from Germany, and you can use it anywhere.

It's a large language model, of course. And you can see how it works right here. Also, interestingly, see, this is in the EU, where they're a bit more careful about stuff, open source models and that sort of thing. So it's designed to be a lot more compliant. Native German and English. It's actually specialized in German on the in-house language data. The model understands German and English when it's running. I wouldn't expect it to run in any other languages. I think it's just running those. And it's also trained, as you can see right here, and built for high-impact use, according to their own words. So essentially, you can see how it performs here. This is pretty interesting. So this is the German benchmark's average score, unweighted average score, across German post-training benchmarks. And if we have a look here, we've got Nematron, we have Gemma 4, we have Quen 3.6b, we have GPT-OSS, and then we have Colibri. And if you look over here, you can see the Colibri is actually scoring highest on that benchmark according to this. Now, also for using tools and agents, this benchmark basically measures how well the model handles multi-step conversations and tool calls as well, which is something, for example, something like maybe you're looking at Hermes agent, or these sort of AI agents, this sort of tests how well it performs. Now, it's tested against Quen, Mistral, Nemetron, and then you can see Colibri over here. So we can get a better look at this.

Yeah. So if we have a look here, you see it's performing pretty well. Mistral actually comes out the worst on this Mistral small four. Then we have Quen 3.6, we have Nemetron 3 Super from Nvidia, Quen is obviously from Alibaba, and then we have CodeLibri over here from Germany. And it's performing pretty well, isn't it? On the benchmarks, of course, always test this stuff out yourself. So you can see here as well, it's a mixture of experts model, 78 billion parameters, 3 billion active, works in German and English. This is pretty good for a local model, a million token context window, pretty good right there. And it's best for agentic tool calling. It will be very interesting to try this out with something like Hermes agent. So you can get the open weights directly from Hugging Face, as you can see here, is designed for text generation. It's not going to generate images and follow that. And let's see if there's anything else we need to know about this.

Knowledge cutoff as well. So it's trained since June 2026, which is pretty useful. Obviously, if you're running it locally and you're not connecting it to the web, you want something that's got access to recent reasoning. And I just dropped 3rd of October. So if you're looking for a free agent, what you could do is you could actually set this up with Hermes agent that would be pretty interesting. Actually, I can see Vasif asked the same question. Watching live, he said, I need a coding agent for free. So I think you could use something like this model with Hermes agent. And if you couldn't, then the best local model that I've seen, bear in mind, I've run a Mac studio, so I don't really run much local AI. But on my Mac studio, the best local model I've tested for Hermes is LFM 2.5 26B.