Mistral's 124B multimodal model with frontier-level image understanding and 69.4% MathVista accuracy.