Mistral Opens Weights for Large 4, Delivering Sovereign AI Without Cloud Dependency

Mistral AI SAS unveiled the open-weight release of Mistral Large 4, its most advanced language model yet, which debuted in preview form three weeks prior. Carrying the nickname "Le Chonk," the model features 1 trillion total parameters with 52 billion active parameters and operates natively across multiple modalities. The system has undergone an additional three weeks of reinforcement learning and real-world customer validation since its initial preview. According to the company, ML4 ranks among the world's leading open-source agentic coding models, delivering competitive performance while consuming significantly fewer tokens and incorporating advanced security defenses.

Pierre Stock, Vice President of Science, stated: "ML4 OSS showcases our science team's ability to innovate at speed. Building on the ML4 Preview, the open-weight version continues reinforcement learning training and delivers substantial gains with no signs of saturation." Mistral trained the model across 3,800 Nvidia Corp. Grace Blackwell graphics processing units housed in European data centers, enabling the company to achieve precision in tool use and coding tasks through complete infrastructure control and hardware-specific optimization during development.

The open-weight approach allows organizations to download and operate the model on proprietary hardware within their own network boundaries, processing confidential information without transmitting it to cloud services or third parties. This capability positions Mistral's Western-developed system as a counterweight to numerous large-scale Chinese coding models currently available, including the GLM family from Beijing Zhipu Huazhang Technology Co. Ltd. and the Kimi family from Beijing Moonshot AI Technology Co. Ltd.

Arthur Mensch, co-founder and Chief Executive, remarked: "Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors. Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion." The training methodology distinguishes ML4 from competing frontier systems, with performance metrics showing it rivals top-tier alternatives such as the Chinese-developed GLM-5.3 on the Artificial Analysis Coding Agent Index while requiring 47% fewer tokens per task on average.

Testing on Terminal-Bench 4.0, which evaluates agentic coding performance in terminal environments, demonstrated ML4's advancement from 28% to 41% accuracy, matching GLM-5.3 as the highest-performing coding agent model. Mistral incorporated three configurable reasoning modes—low, medium and high—enabling teams to calibrate reasoning depth according to specific task requirements without modifying prompts, a technique that often proves less reliable. This design grants developers and enterprise customers granular control over token consumption per individual task.

The ML4 release marks the beginning of a new generation spanning Mistral's complete model portfolio across all modalities. The announcement follows Mistral's recent introduction of Voxtral 3, a suite of voice models and voice agents. The company has signaled plans to introduce additional models from this expanded family over the coming months.