Mistral AI releases Mistral Large 4 trillion-param model

French AI company Mistral AI launched Mistral Large 4 ("Le Chonk") as a public preview on October 6, 2026, a 1 trillion-parameter multimodal mixture-of-experts model with 49 billion active parameters. The model represents Europe's push for open-weight AI alternatives to American and Chinese systems, with open-weight release planned for late October and the capability to compete with GPT-5 and Claude Opus 4 series on coding and reasoning benchmarks.

Mistral Large 4, internally nicknamed 'Le Chonk', marks a significant milestone for European AI development and open-source advocates. With 1.05 trillion total parameters and 49 billion active per token, the model is natively multimodal, accepting both text and image inputs while responding in text. The architecture uses a granular mixture-of-experts approach, activating only about 4.7% of its weights per token-a design that allows a trillion-scale model to operate at mid-tier pricing.

Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, Large 4 was trained on data spanning more than 160 languages, including every official EU language. The model currently operates through the Mistral Studio API and official API at $1.36 per million input tokens and $4.18 per million output tokens. Early benchmarks indicate improvements across multiple domains, particularly in complex reasoning tasks and code generation, with the company claiming top-five status on the Artificial Analysis Cyber Index and superiority over GPT-6 Astra on certain legal and finance tasks.

Mistral positioned Large 4 as a 'third way' in AI between closed American models and open-source models often developed in China. The model supports structured outputs, function calling, document Q&A, batch processing, and agentic capabilities through its API. Mistral plans to release the model weights by the end of October 2026, allowing developers to download and run it locally. A notable feature includes two-tier access: the public API maintains standard safeguards, while a private API for 'privileged partners' in cybersecurity and state authorities offers reduced moderation and broader offensive and defensive capabilities, currently undergoing red-teaming exercises before weights release.