SIMA 2 is a research agent from DeepMind that integrates Gemini model capabilities to operate in complex 3D virtual environments. It can follow human-language instructions, engage in dialog, plan toward goals, and control in-game inputs in real time while reading pixel and UI observations. SIMA 2 generalizes skills across novel games, accepts multimodal inputs, and learns from human guidance, enabling interactive human-in-the-loop training and open-ended skill acquisition. Its design advances general, helpful agents for research into reasoning, learning, and interactive behavior in simulated worlds.