Mistral AI says its Mistral Large 4 model wrote, validated and visualized a self-consistent field solver from a short natural-language prompt. The model compared its results with existing tools, then generated an interactive 3D view showing how orbitals behave as a chemical bond forms. Mistral Large 4 has approximately 1 trillion parameters, including 52 billion active parameters per token, a 1 million-token context window and a 1.6 billion-parameter visual encoder. The model was trained on 3,800 NVIDIA Grace Blackwell GPUs before its weights were opened on October 29, 2026, enabling researchers to independently test its capabilities despite substantial hardware requirements.
Source: Read the original article


