Google has rolled out a significant update to its Gemini 3 Flash AI model, introducing a feature called Agentic Vision. This new capability enhances the model”s proficiency in analyzing intricate images, allowing it to detect fine details such as serial numbers and text within complex diagrams.
With Agentic Vision, Gemini 3 Flash has adopted a more interactive approach to image analysis. The model is designed to “think and act” by executing code, which improves its overall functionality. As part of its future developments, Google aims to incorporate capabilities like image web search into this model.
The introduction of Agentic Vision marks a pivotal evolution in agentic models, which not only provide answers but also perform intermediate actions to refine results. Google described this feature as a breakthrough for “frontier AI models,” enabling a Think, Act, Observe loop in image comprehension tasks. This process involves analyzing a user prompt and the initial image, generating and executing Python code to manipulate the image, and then reassessing the modified visuals for more accurate outcomes.
The key functionalities of Agentic Vision include:
- Planning: Developing a systematic strategy for image analysis.
- Zooming: Automatically magnifying small elements for better clarity.
- Annotations: Marking up images to enhance the model”s reasoning.
- Visual math and charting: Interpreting dense tables and using Python for visual representation of results.
This feature is currently accessible through the API and is being showcased in Google AI Studio. For instance, an AI-powered platform named PlanCheckSolver.com, which reviews construction plans, reported a 5% increase in accuracy after implementing Gemini 3 Flash”s code execution capabilities. The model is able to generate Python code to isolate specific elements of an image and reintroduce them into the analysis for compliance verification.
Another example includes the Gemini app, where the model accurately counted fingers on a hand by using Python to create bounding boxes and numeric labels on each finger, resulting in a “visual draft” that minimized errors.
Moreover, Agentic Vision facilitates processing complex tables and constructing charts with tools like Matplotlib. By delegating calculations to a deterministic Python environment, the model moves beyond probabilistic assessments.
Google has indicated that this is merely the initial phase of Agentic Vision”s development. The tech giant is exploring ways to enhance functionality, such as allowing image rotations and advanced calculations to occur without explicit user prompts. Future updates may also include the integration of new tools like web search and reverse image search, with plans to expand these capabilities to other sizes of the Gemini model beyond just Flash.
In a notable collaboration, Apple has confirmed a multi-year partnership with Google, aiming to build the next generation of Apple Foundation Models on Gemini models and Google”s cloud infrastructure. This partnership will enhance Apple Intelligence features, including a revamped and personalized version of Siri.
