Mistral Large 4
7 storiesMistral Large 4 is a trillion-parameter open-weight multimodal model from the French lab Mistral AI, built as a fine-grained Mixture-of-Experts with 52B active parameters, 1.05T total parameters and a 1.6B vision encoder. Recent coverage focused on its October 2026 preview and weight release, claims of using 47% fewer tokens per task than GLM-5.3, a partnership with Nvidia on physics AI for industrial design, and a French tax authority contract after a major hack.
Related topics
After a large-scale hack in June and July, France's tax authority accelerated awarding a contract to domestic AI firm Mistral. Mistral will initially be deployed at DGFiP to strengthen protection and will compare social media accounts and shopping sites with tax returns to spot anomalies; the Mistral contract is just €6 million of a €200 million cybersecurity budget increase.
Mistral released the weights of its trillion-parameter Mistral Large 4, pitched at sovereign cyber defense. It claims performance comparable to GLM-5.3 while using 47% fewer tokens per task, and says ML4 tops other open models on Artificial Analysis' Coding Agent Index.
Mistral AI released Mistral Large 4 in public preview, an open-weight, general-purpose multimodal model with a fine-grained Mixture-of-Experts architecture: 52B active parameters, 1.05T total parameters, and a 1.6B vision encoder.
Mistral and Nvidia are combining Mistral Large 4 and Mistral's physics AI with CUDA-X, Warp and Omniverse to link design, simulation and AI training for aerospace, automotive and semiconductor industries; they also plan public benchmarks and open datasets. The roughly 1-trillion-parameter Mistral Large 4 was trained on 3,800 Grace Blackwell GPUs in Europe and ships as open weights in late October.
Open-source model GLM 5.3 Flash ranks first on the Artificial Analysis Cyber Index, surpassing every Anthropic model. Two open models top the leaderboard, showing open-source competitiveness in cyber capabilities.
Mistral's largest open-weight frontier model, a 1T-parameter natively multimodal MoE for developers.
French lab Mistral AI released Mistral Large 4, a 1-trillion-parameter multimodal model nicknamed Le Chonk, accessible only via a guarded public endpoint with open weights promised in three weeks after safety testing. Trained on just 4,000 NVIDIA GPUs, using two to three times less compute than Chinese rivals, it targets use cases like cybersecurity.
That is everything