NEW Mistral Le Chonk DESTROYS Kimi K3?

Julian Goldie (0:00)

Mistral just dropped their biggest AR model ever, Mistral Large 4, and I have to be honest with you about it. Everyone's calling this Europe's answer to the top AI models, a giant one trillion parameter brain. So I put it to the test head to head against the best open models in the world. What happened next genuinely shocked me. So in this video, I'll show you exactly how it performed, what it's actually good at, and the one thing it beats almost every other model on, which I'll reveal later, because if you pick the wrong AI model right now, you're leaving massive results on the table. By the end of this video, you'll know exactly which model to use, which is the best open model to use, and how to get the best work done faster with open source models. Let's get straight into it. So we have a brand new model from Mistral called Mistral Large 4, aka Le Chonk. And I think that's a reference to Le Chaton Grosse, which was an Internet meme around AI agents. If you saw that a few months ago. Anyway, this is a 1 trillion parameter natively multimodal mixer of experts model.

It's stated as the best open-weights model from the US or Europe on aggregated benchmarks. Let's see on that in a second. And it's designed to have a state of the art on critical workloads, including cyber defense manufacturing and surpasses frontier models on visual grounding. Again, let's check that in a second. This is made in Europe, comes from Mistral, which is a French company. And you see the details right here of Le Chonk. Now, this is available as a public preview. So you can see it's available in public preview right now. And it's an open-weight model. Though weights are not open currently, but you can, will get access to that later. That's coming. And there's two ways you can access it right now. So you can get the API with Mistral. If you sign in to their studio, you can get the API and that's available on a pay plan. The other option is that you can actually get access on OpenRouter.

So on OpenRouter, you can get access to this.

Again, this is not like a frontier model, but it's designed to be one of the best open-waist models. We'll come on to that in a second, see how it performs. So if we have a look at the Artificial Analysis Index here, you can actually see that on their scores, it's coming in at a total score of 38, which is still below DeepSeq, GLM 5.3 Flash, Kimi K3 Max, GLM 5.3 Max, and MiMoV 2.6 Pro. So if you look at all of the OpenWeights models here, it's still not right at the top of the leaderboard here, and Chinese models are still outperforming it on Artificial Analysis Index scores.

We can compare it on all of the metrics here, which is pretty useful. So if you want to go super in-depth on that, you can see how it performs versus all of the other open models as well. Now, it's designed to be Frontier Performance, so it's a trillion parameter model, 52 billion active parameters, their largest and most capable model to date. So this is like the frontier model for Mistral so far. And if you look at, for example, like DeepSWE, Kimi K3 still outperforming it, still way ahead.

If we have a look at coding, GLM 5.3 still beating it.

Cyber, it seems to be leading out of all of the models here, but it's on par, right? So it's on par with GLM 5.3. That seems to be the only thing that it's outperforming everyone on. If we look at agentic behavior here, it's just slightly below GLM 5.3. And grounding scores 42 versus GPT-6 Astra at 41.5. It was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Europe.

And one of the things that it probably performs the best on is cybersecurity out of everything. When I'm looking at the benchmarks here. Now you can try the preview API, or you can just go to open router over here. You can see the costs. So 0.68 dollars input, $2.09 output per million tokens, 500k context window, just released on October the 6th. And you can start building with it over here. OpenWay is not released yet, but they are supposed to be coming soon. You might also wonder what are the best apps to use it in. So far, most people are using it like KiloCode, Pi, Hermes Agent, Claude Code, and some others. Let's see how it performs over here. So I'm going to say, for example, we'll switch on the tools, create a beautiful website for the AI Profit Boarding Community, Community About AI Automation. So we can see how it performs there. And what we'll actually do is we'll compare it against Kimi K3 and see how they perform on both speed and also performance. Now, if we have a look here, we've already got the output from Kimi K3. So that's completed. If you look at Mistral, number one, Mistral seems to still be thinking. So it's still in thinking mode here. And it's got nowhere near creating the actual HTML. So it seems a lot slower as well side by side. GLM 5.3 and Kimi K3 were both pretty good models.