This text announces Gemini 3 Pro, a multimodal AI model from Google DeepMind, presented as a significant advancement in visual and spatial reasoning. The model demonstrates state-of-the-art performance on benchmarks for document, spatial, screen, and video understanding. Key capabilities highlighted include "derendering" complex documents into structured code, outputting precise pixel coordinates for spatial tasks, robustly understanding and interacting with on-screen user interfaces, and processing high-frame-rate video for detailed causal analysis. The announcement also covers real-world applications in education, medicine, law, and finance, and introduces a new media resolution control for developers to balance performance and cost.
Gemini 3 Pro: A New Multimodal AI with Advanced Visual Reasoning
Overall Capability: Gemini 3 Pro is described as a major leap from simple recognition to complex visual and spatial reasoning, setting new performance records on vision benchmarks like MMMU Pro and Video MMMU.
Document Understanding: The model excels at parsing unstructured documents. It can "derender" visual documents (e.g., historical logs, handwritten math) into structured code like HTML or LaTeX and perform multi-step reasoning across long, complex reports, supposedly outperforming human baselines on some benchmarks.
Spatial Understanding: It can identify specific locations in images by outputting pixel-precise coordinates ("pointing capability"). This is intended for applications in robotics (generating spatially grounded plans) and augmented reality (e.g., pointing to a specific screw based on a manual).
Screen Understanding: The model applies its spatial capabilities to understand desktop and mobile UI screens, enabling the creation of robust agents that can automate repetitive digital tasks, such as creating a pivot table in a spreadsheet.
Video Understanding: Gemini 3 Pro features significantly improved video analysis. It can process video at high frame rates (e.g., 10 FPS) to capture rapid action, uses a "thinking mode" to understand cause-and-effect relationships over time, and can convert information from long videos into functional code.
Real-World Applications: The model is positioned for use in education (solving diagram-based science and math problems), medicine (analyzing biomedical imagery), and law/finance (processing complex reports and contracts).
Developer Controls: A new media_resolution parameter gives developers control over visual token usage, allowing them to balance image fidelity against cost and latency for different tasks.