Mistral releases new multimodal model with one trillion parameters

French artificial intelligence laboratory Mistral AI has introduced a new multimodal model called Mistral Large 4, or ML4. The system is designed to compete with both American and Chinese alternatives while offering a distinct European option in the artificial intelligence landscape.

Architecture and access

Nicknamed Le Chonk due to its massive scale, the model contains one trillion parameters — the variables a neural network adjusts during training to recognize patterns. While the system is currently accessible only through a public guardrail endpoint, Mistral plans to release the model weights in three weeks after completing safety testing.

Pierre Stock, vice president of science at Mistral, explained that the temporary restriction allows the lab to work with trusted partners and governments. This interim step ensures the future open-source weights can support defensive use cases rather than malicious applications, addressing growing security concerns among enterprise customers and public institutions.

Compute efficiency

An unusual aspect of the model's development is its training infrastructure. Mistral trained the system using only 4,000 Nvidia graphics processing units — the specialized microchips used to accelerate machine learning workloads. According to the lab, this compute expenditure is significantly lower than that of closed-source competitors and two to three times less than Chinese rivals.

Target applications

Mistral optimized the model for specific professional domains where multimodal capabilities add practical value. These target use cases include cybersecurity, finance, and chip design. The latter application aligns closely with two of the lab's major financial backers, Dutch semiconductor equipment maker ASML and Samsung, which led a recent funding round valuing the company at approximately €21 billion, or roughly $24.39 billion.