Mistral Open-Sources 'Le Chonk': a 1.05T MoE Topping Open Coding

Europe's most important model lab has finally handed over its biggest model. Mistral has released the weights of Mistral Large 4, nicknamed "Le Chonk" internally: a sparse mixture-of-experts model with 1.05 trillion total parameters, about 52 billion active per token, a 1-million-token context window and native image input.

Earlier this month it was a public preview. What ships now is an improved version, and Mistral is leaning hard on reasoning efficiency: the same quality at a lower inference cost. By the company's account it is the top open model on Cognition's FrontierCode coding benchmark, and the best open-weight option on defensive cybersecurity, finance and manufacturing tasks.

A trillion parameters, used a sliver at a time

Architecturally, Le Chonk follows what has become the default recipe for frontier open models: a huge expert pool plus sparse attention. It has 72 decoder layers; a router picks 4 of 128 experts per token, alongside one always-on shared expert. Attention pairs multi-head latent attention with a DeepSeek-style sparse indexer that lets each query look at no more than 2,048 earlier positions. A 1.6-billion-parameter Pixtral-style vision encoder handles images of varying size and aspect ratio.

This is close to the industry's consensus answer, from DeepSeek to GLM: cut cost with two kinds of sparsity at once, one over experts and one over context. Mistral didn't reinvent the architecture. It spent its effort on data and vertical skills, which is a pragmatic call for a company that needs to turn models into revenue.

The model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers. "Trained and hosted in Europe" is itself the pitch. As sovereign AI turns into a procurement requirement for governments and large enterprises, that line may move more deals than a few benchmark points.

Open, with one door kept shut

The license is Mistral's new Ecosystem Open-Weight License, a modified MIT. Anyone can use, modify, fine-tune and commercialize the model for free, with one exception: if you want to sell it as a model-as-a-service API, you need a separate commercial license from Mistral.

That clause is plainly aimed at hyperscalers and inference hosts. Mistral has watched others package its open models and capture the revenue; this is its attempt to balance community goodwill with a commercial moat. For most companies deploying the model for their own use, it changes nothing.

The hardware bar is real, though. The official FP8 checkpoint wants eight B300s; the NVFP4 build needs four GB300s and runs only on Blackwell. Unsloth and llama.cpp have already posted GGUFs, but no amount of quantization makes a trillion-parameter model a local toy. For most developers the realistic entry point is the API, priced at $0.70 per million input tokens, $0.07 cached and $2.10 output, which is aggressive for this weight class.

Europe's flagship finally carries weight

For two years the open frontier has largely belonged to Chinese labs, American big tech has cooled on open releases since Llama, and Mistral was widely seen as having fallen behind in the race for the strongest model. Le Chonk gives Europe a trillion-scale open model that can stand next to DeepSeek, Qwen and Kimi.

Open-model leadership is still not the closed frontier, and a FrontierCode lead may not last long, with Qwen4 reportedly due next week. But for companies that need data to stay inside the EU, and for teams that want to own their model for coding and security work, Mistral has created an option that didn't exist before.