Gemini 3 Flash introduces Agentic Vision, transforming static image analysis into an active, iterative process. Unlike traditional frontier AI models that rely on a single glance—often missing fine details like microchip serial numbers or distant signs—Agentic Vision employs a “Think, Act, Observe” loop. It first analyzes the query and image to plan multi-step actions; then executes Python code to zoom, crop, enhance, or inspect regions; and finally observes results to ground answers in visual evidence. Code execution, the first supported tool, boosts vision benchmark performance by 5–10%. This agentic approach enables precise, evidence-based visual reasoning—marking a significant leap in AI’s ability to dynamically investigate and understand complex imagery.

本专栏通过快照技术转载，仅保留核心内容

内容中包含的图片若涉及版权问题，请及时与我们联系删除