European AI flag bearer Mistral's new open weights model is 'Le Chonk'

Another week brings yet another open weights model and this one is huge — or at least it is for Europe's AI flag bearer Mistral, which has taken to calling the 1 trillion-parameter large language model (LLM) Le Chonk.

Its product name is Mistral Large 4, which is a lot less fun than the cat-meme-inspired "very official" moniker Le Chonk. Nonetheless, the LLM is the French model-dev's biggest open weights release ever.

At a trillion parameters, Le Chonk is squarely in frontier territory, and like most modern frontier models, it is multimodal, reasoning, and employs a mixture of experts (MoE) architecture with about 49 billion active parameters to keep serving costs low.

REG AD

Mistral didn't offer any guidance on hardware requirements in its release blog, but the model is still small enough to run reasonably on 8-way GPU boxes like Nvidia's HGX B300 or AMD's MI355X.

REG AD

We know far more about the hardware the model was trained on. According to Mistral, this Chonker was trained using roughly 3,800 Grace Blackwell GPUs — or around 52 NVL72 racks — in Mistral's own datacenters located in Europe on a corpus spanning more than 160 languages using a combination of supervised pre-training and reinforcement learning.

Sovereign AI remains a major focus for Europe and Mistral is clearly cashing in on demand for frontier class models trained and run within Europe's borders.

The model itself is currently in preview, but you won't have to wait long for its weights to drop. Mistral expects them to be live on Hugging Face and other model repos within the month.

Early testing by independent benchmarking firm Artificial Analysis puts Le Chonk at a major disadvantage relative to OpenAI or Anthropic's flagship models. In the AI guru's intelligence leaderboard, Mistral Large 4 preview sits between DeepSeek V4.1 Flash and OpenAI's entry-level GPT6 Luna models.

Having said that, Mistral's latest model outshines the United States' most capable open weights model, Thinking Machines Lab's Inkling, by a significant margin. So when it comes to non-Chinese LLMs you can actually download, Mistral does quite well.

As you'd expect, Mistral's own benchmarks tell a far rosier story. Across coding, agentic, finance, and legal benchmarks, the model build shows its latest LLM trading blows with some of the top Chinese offerings from Alibaba, Moonshot, Z.AI, and DeepSeek. You can find their charts here, but as always take them with a grain of salt.

One area Mistral is particularly proud of is Le Chonk's performance in cybersecurity workloads. The company claims the model is one of the strongest performers in cybersecurity benchmarks like Artificial Analysis' new Cyber index.

This is a dig at proprietary American models from the likes of OpenAI and Anthropic, which have a tendency to refuse cybersecurity-related tasks that could be perceived as adversarial.

REG AD

As we previously reported, this is potentially problematic for customers trying to use AI models to red-team their own systems in order to identify vulnerabilities before someone else finds them.

"Provider-level refusals can block legitimate vulnerability research and incident response, and where losing access to a capability mid-incident can itself become a critical security risk. ML4 pairs top-tier cyber performance with open weights and self-deployment, giving organizations both the capability and the autonomy to run advanced security work under their own policies," Mistral wrote in a blog post.

According to Mistral, Le Chonk is only the first in a series of new models made possible by its €3 billion (about $3.4 billion) Series D funding round last month. The company already expects the model's performance to improve meaningfully as it continues applying reinforcement learning, which suggests a more potent Mistral Large 4.1 could be around the corner.

If you'd like to give the model a spin before its weights drop later this month, it's currently available via Mistral's API. ®