We present MedMax, a large-scale multimodal biomedical instruction-tuning dataset for mixed-modal foundation models, and show that our data achieves superior performance than GPT-4o on diverse biomedical tasks.