Mistral Launches 1 Trillion-Parameter ML4, Promising Open Weights in 3 Weeks
Updated
Updated · TechCrunch · Oct 6
Mistral Launches 1 Trillion-Parameter ML4, Promising Open Weights in 3 Weeks
3 articles · Updated · TechCrunch · Oct 6
Summary
Mistral on Tuesday released Mistral Large 4, a 1 trillion-parameter multimodal model now available only through a guarded public endpoint, with open weights slated for release after three weeks of safety testing.
The French lab is pitching ML4 as a European "third way" between closed U.S. systems and often-China-built open models, aiming to serve enterprises and institutions worried about security while preserving auditability.
Pierre Stock said ML4 was trained entirely on Mistral's own compute using 4,000 Nvidia GPUs—two to three times fewer than Chinese rivals and well below closed-model competitors.
Mistral is targeting cybersecurity, finance and chip design, areas tied to customers and backers including ASML and Samsung, as it tries to prove it remains a frontier lab rather than just an inference provider.
Benchmark results are still pending, but Mistral says focused training could make ML4 best in class among open-weight models outside China and competitive with closed models in selected tasks.