On October 6, 2026, the French AI company Mistral AI released a public preview of its new large language model, Mistral Large 4. Internally, the model goes by the nickname "Le Chonk."
The announcement
According to the official announcement, Mistral Large 4 is a Mixture of Experts (MoE) model with 1 trillion total parameters, of which 49 billion are active for each token processed (active parameters). It is a multimodal model that handles both text and images, and a single model offers both a mode that answers quickly and a mode that takes time to reason through problems.
It was trained on 3,800 NVIDIA Grace Blackwell GPUs in data centers that Mistral operates itself in Europe. The company says it supports more than 160 languages, including all official languages of the EU.
Key numbers
These are the main figures Mistral published.
- 61.7% on DeepSWE v1.1, a software development evaluation
- 28.3% on Terminal-Bench 4.0, an agentic coding evaluation
- 59.9% on AutomationBench, a business automation evaluation
- 93% on Cybench, a set of 40 security competition challenges
- API pricing of $1.36 per million input tokens and $4.18 per million output tokens
All benchmark figures come from Mistral's own announcement. Mistral says it significantly outperforms any open-weight model developed in the US or Europe.
Availability and what comes next
The model is currently available as a preview through the API of Mistral Studio, the company's platform for developers (model ID: mistral-large-4-0). Mistral plans to publish the model weights by the end of October, after which anyone will be able to download them and run the model on their own servers.
Mistral notes that the preview is still partway through tuning with reinforcement learning, so its quality may differ from the final release. Because the model is highly capable in cybersecurity, the company is also working with security firms, public institutions and others to test it (red teaming).
CoAI's take
High-performance open-weight models from European companies that can run in your own environment are a significant option for businesses that do not want their data to leave the company. That said, running a model with 1 trillion parameters in-house requires substantial computing resources. A realistic approach is to try its performance through the API first, then check the license terms and hardware requirements once the weights are published at the end of the month.
