Introducing Mistral Large 4

Mistral has announced a public preview of its new large multimodal model, Mistral Large 4 (ML4), described as a 1‑trillion‑parameter model with 52 billion active parameters. The preview is available on Mistral Studio, with full weights scheduled for release by the end of the month. ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters and is positioned as a state‑of‑the‑art open‑weight model for cybersecurity, coding, and multimodal understanding. According to the company, ML4 achieves performance competitive with the strongest open‑source models globally and significantly outperforms any open‑weight model developed in the US or Europe. In cybersecurity, the model ranks among the top five on the Artificial Analysis Cyber Index and leads open‑weight models outside China, scoring 82% on a vulnerability‑reproduction test and 93% on the Cybench exercise. In coding, ML4 achieves high scores on DeepSWE v1.1, SWE‑Atlas‑QnA, and Terminal‑Bench 4, and ranks second in a blind human evaluation of coding quality. The model also excels in visual grounding, surpassing GPT‑6‑Astra on Dense 200, and demonstrates strong scientific capabilities, achieving state‑of‑the‑art results on SciCode‑Verified. Mistral emphasizes that the model’s open weights and self‑deployment enable organizations to run advanced security and research workloads under their own policies, addressing concerns about provider‑level refusals and mid‑incident access loss. The company plans to share further architectural details and benchmarks as it prepares to release the weights.