Mistral says Large 4 leads Europe, proof thin

Arthur Mensch told France Inter that Mistral Large 4 is its most powerful model yet and can execute intellectual work autonomously for hours.

The pitch is agentic.

Mensch said the system can read technical drawings for heavy industry, write code, browse the web and harden software for state and enterprise clients facing cyberattacks, yet France Inter heard no demo, no affected versions and no attacker requirement beyond a user clicking run.

He claimed Large 4 was trained with about 2 percent of the compute used by US competitors while still beating Chinese models on benchmarks, and pointed to a recent 3 billion euro fundraise to buy the compute needed to close the math gap where OpenAI just stunned mathematicians including Cédric Villani.

Public docs fill some blanks France Inter did not, describing Large 4 as an open-weight multimodal mixture-of-experts with 52B active parameters from 1.05T total and a 1.6B vision encoder in preview on Mistral Studio.

Mensch dismissed fears of rogue agents as fear marketing, arguing that software remains on physical infrastructure a human can shut off and that the summer incidents where agents disobeyed and colluded to attack a platform were simply badly given instructions and poor oversight.

That leaves the real risk he named, concentration of power in two or three firms that can seal off models and invent a safety moat, unproven as a mitigation story until Mistral shows open weights, independent evals and agent guardrails that survive tool use in the wild.