Mistral Large 4 "Le Chonk" by the numbers

A 1-trillion-parameter model called Le Chonk just landed. It’s the first open-weights model from Europe or the US to reach that scale. The jump from Mistral Large 3’s 675 billion parameters is massive. Yet despite the size, it’s still eight points behind Xiaomi’s MiMo-V2.6-Pro on the Artificial Analysis Intelligence Index. The numbers show parameter count alone isn’t closing the gap with Chinese rivals.

What 1 trillion parameters actually buys you

Mistral Large 4 packs 1 trillion parameters. That’s 325 billion more than its predecessor. For context, Mistral’s own Large 3 topped out at 675 billion. The jump is real, but benchmarks tell a more complicated story.

On the Artificial Analysis Intelligence Index, Le Chonk scores 38. That’s ahead of other Western open-weights models, but still eight points behind Xiaomi’s MiMo-V2.6-Pro, which scores 46. Mistral calls this the “strongest open-weights model from the US or Europe,” and the claim holds up, but only because the competition is weak. The gap with China is now measured more clearly.

What’s missing is inference speed. Mistral hasn’t disclosed hardware requirements or tokens-per-second benchmarks. A trillion parameters looks impressive, but if it takes a rack of H100s to run a single query, the practical advantage disappears fast.

Multimodality: impressive on paper, unproven in practice

Le Chonk handles images, text, and code out of the box. Internal evaluations used expert annotators in coding, CAD, finance, math, and physics, comparing it directly to GLM-5.3. That’s smart, GLM-5.3 is one of the few Chinese models with similar multimodal capabilities, so the comparison is fair. But here’s the catch: Mistral hasn’t released the raw scores or methodology. Without public benchmarks, especially in real-time multimodal tasks like video or audio, it’s hard to separate hype from substance.

The API opened on October 6, and the weights drop at the end of October. That timeline is aggressive, but it’s also revealing. Mistral is pushing this as a production-ready model, not a research experiment. The question is whether the multimodal capabilities are robust enough for enterprise use, or if they’re still in the “demo-ware” phase. Early adopters will find out soon.

Cost and pricing: the brutal math

Mistral Large 4 costs $1.13 per task on the Artificial Analysis Intelligence Index. That’s 38 times more than Mistral Large 3’s $0.03 per task. Token pricing is $0.68 per million for input and $2.09 per million for output, which undercuts competitors. On paper, that looks competitive. In reality, it’s more complicated.

The per-task cost jump is steep. Mistral is betting that the quality uplift justifies the price, but for most users, the math won’t add up. If you’re running a high-volume chatbot, the token pricing might work. If you’re doing anything more complex, multimodal reasoning, long-context tasks, the cost could spiral. The pricing strategy feels like a hedge: Mistral is positioning Le Chonk as a premium model, but the numbers suggest it’s still not cost-competitive with Chinese alternatives.

Infrastructure and sovereignty: built in Europe, for Europe

Mistral built Le Chonk entirely in Europe, and it’s deployable on Mistral Cloud from European data centres. That’s a big deal for EU customers with data-localization requirements. For companies in finance, healthcare, or government, this is a compliance necessity. Mistral is playing the sovereignty card hard, and it’s working. US cloud providers can’t match this, and Chinese models come with their own geopolitical risks.

But there’s a catch. Mistral Cloud is a closed platform. If you want to self-host Le Chonk, you’re on your own. The open-weights release at the end of October will help, but without disclosed training FLOPs or dataset composition, it’s unclear how reproducible this model really is. For now, the sovereignty play is real, but it’s limited to Mistral’s own infrastructure.

Architecture: granular MoE, but no transparency

Le Chonk uses a granular Mixture-of-Experts design. That’s expected, MoE is the go-to for scaling beyond a few hundred billion parameters, but Mistral hasn’t shared the details. How many experts? How are they routed? What’s the active parameter count per token? Without answers, it’s hard to assess the efficiency of the architecture.

Mistral calls Le Chonk an “open-weights model,” but that’s a loose term. Open weights don’t mean open source. You can download the weights, but you can’t inspect the training process, dataset, or FLOPs. That’s a problem for reproducibility. If Mistral wants to compete with truly open models like Llama, they’ll need to open up more than just the weights.

Licensing: the MIT shuffle continues

Mistral’s licensing history is inconsistent. Large 3 and Small 4 shipped under Apache 2.0, which is as permissive as it gets. Medium 3.5 used a modified MIT license with added restrictions. For Le Chonk, Mistral hasn’t specified the license yet. That’s a red flag. If they stick with Apache 2.0, fine. If they introduce another custom license, it’ll limit adoption.

The ambiguity matters because licensing is a moat. Chinese models like MiMo-V2.6-Pro have clear, permissive licenses. If Mistral adds restrictions, they’re handing ground to competitors. For now, the lack of clarity is the biggest barrier to adoption.

Where Le Chonk still trails Chinese models

The eight-point deficit on the Artificial Analysis Intelligence Index shows a real gap in capability. MiMo-V2.6-Pro scores 46, Le Chonk scores 38. That’s the difference between a model that can handle complex multimodal reasoning and one that’s still catching up.

The bigger issue is inference speed. Mistral hasn’t disclosed hardware requirements, but a trillion parameters likely means high latency and high cost. Chinese models are optimized for efficiency, not just scale. If Le Chonk can’t match their speed or cost, the parameter count won’t matter.

Then there’s real-time multimodality. MiMo-V2.6-Pro handles video and audio natively. Le Chonk doesn’t. That’s a dealbreaker for applications like live transcription, video analysis, or robotics. Mistral is strong on static multimodality, but they’re still behind on dynamic tasks.

Early adopter feedback: silence speaks volumes

Le Chonk has been in public preview since October 6, but there’s no aggregated user feedback yet. No reviews, no third-party benchmarks, no real-world usage data. That’s a problem. Mistral is pushing this as a production-ready model, but without early adopter validation, it’s still a black box.

The silence is telling. Either users are under NDA, or they’re not seeing the results Mistral claims. Either way, the lack of public feedback makes it hard to assess the model’s real-world performance.

What this means for open-weight AI

Mistral calls Le Chonk the “strongest open-weights model from the US or Europe.” The claim holds up on benchmarks, but it’s a low bar. The real competition is in China, and Mistral isn’t winning there. The bigger question is what this means for open-weight AI as a category.

The US and EU are pushing for open-weight leadership, but Le Chonk shows the limits of that approach. Scale alone isn’t enough. Efficiency, licensing, and real-world performance matter just as much. If Mistral wants to compete with Chinese models, they’ll need to open up more than just the weights, they’ll need to open up the entire pipeline.

For now, Le Chonk is a step forward, but it’s not a leap. The numbers tell the story: 1 trillion parameters, 38 on the index, $1.13 per task. That’s progress, but it’s not dominance. The question is whether Mistral can turn this into a sustainable advantage, or if it’s just another arms race with diminishing returns.

Mistral Large 4: "Le Chonk"