Mistral’s new 1T model aims to leapfrog closed and open rivals

Key Points:

• French AI lab Mistral AI has launched Mistral Large 4 (ML4), a large multimodal AI model with 1 trillion parameters, aiming to compete with American and Chinese AI technologies while offering a "third way" as described by President Macron.
• ML4 is currently accessible only via a public guardrail endpoint, with plans to release open weights in three weeks after safety testing, balancing openness with security concerns to prevent misuse.
• The model was trained using Mistral's own compute resources, specifically 4,000 Nvidia GPUs, which is significantly fewer than competitors in China and closed-source models, highlighting efficiency in training.
• Mistral aims for ML4 to be a top-performing open-weight model globally, with particular strengths in cybersecurity, finance, and chip design, aligning with the interests of major backers like ASML and Samsung.
• The release reflects ongoing tensions between open and closed AI models, with Mistral positioning ML4 as a secure, auditable alternative that addresses enterprise and institutional security concerns.