Mistral Large 4

Mistral AI announced the public preview of its newest model, Mistral Large 4 (ML4), a trillion‑parameter, multimodal system with 49 billion active parameters that can be accessed today via the Mistral Studio API, while the model weights are slated for release at the end of the month. Built from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company’s European data centres, ML4 is positioned as the most capable open‑weight model Mistral has produced, delivering state‑of‑the‑art performance on enterprise‑critical tasks such as cybersecurity, finance, and law, and even surpassing leading closed‑source models in visual‑grounding benchmarks. The preview is being red‑teamed with cybersecurity leaders, vetted partners, and state authorities who receive a version with reduced moderation and expanded cyber capabilities, underscoring Mistral’s focus on giving organisations autonomous, self‑deployed AI that can operate under their own policies and legal frameworks.

The model’s most striking consequence is its demonstrated superiority on security‑focused evaluations, where it ranks among the top five AI systems globally on the Artificial Analysis Cyber Index and leads all open‑weight models developed outside China. In a test requiring the reproduction and patching of a real vulnerability in open‑source software, ML4 achieved an 82 percent success rate, the highest recorded, while closed models such as Claude Opus 5.5 and GPT‑6 Astra refused the task entirely. Similarly, on the Cybench suite of 40 security‑competition challenges, ML4 solved 93 percent of the problems, a score unmatched by comparable open models. These results highlight the practical advantage of an unrestricted model for defenders, who need to verify flaws and craft mitigations without the safety filters that often block such work in proprietary systems, especially as threat actors increasingly jailbreak those same models for offensive purposes.

Beyond cybersecurity, ML4 shows strong capabilities across software engineering benchmarks, scoring 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE‑Atlas‑QnA, and 28.3 percent on Terminal‑Bench 4, and achieving a 49.8 percent combined Coding Agent Index score that outpaces rivals like DeepSeek V4 Pro 0813 and Qwen 3.8 Max. The model was trained on a multilingual dataset covering more than 160 languages, including every official EU language, and was developed in collaboration with leading enterprises across finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping, and the public sector. Mistral plans to use ML4 as a foundation for a new generation of specialised models, and will later disclose additional architectural details, benchmarks, and post‑training methodology. Organizations can already test the preview API, provide feedback, and anticipate the ability to run the model on private cloud or on‑premise infrastructure, ensuring sovereign, auditable AI for mission‑critical operations.

⚡ Effects Interpreter