In this deep dive into Gemini 2.0, Google explains how its new code execution capabilities empower Gemini models with a Python sandbox for on-the-fly computations, data analysis, and visualization. Now available in both Google AI Studio and the Gemini API, the feature lets models run Python code for up to 30 seconds at a time and execute tasks—such as calculating "the sum of the first 50 prime numbers"—with popular libraries like Numpy, Pandas, and Matplotlib. This upgrade not only supports file inputs and complex graph outputs but also paves the way for experimental features like real-time multimodal interactions and integration with tools such as the Multimodal Live API and Grounding with Google Search.

Key Points:

• Gemini 2.0 now features code execution in a Python sandbox available via Google AI Studio and the Gemini API.
• The code execution tool runs code for up to 30 seconds at a time and can be used up to 5 times per session without re-prompting.
• It supports popular libraries such as Numpy, Pandas, and Matplotlib, with more libraries expected in the future.
• New features include file I/O, and support for generating complex visualizations like graphs and charts.
• Practical examples include real-time data analysis with bar chart generation for Tom Cruise movie runtimes and solving optimization problems for traveling salesmen in Spain.
• Experimental capabilities combine code execution with the Multimodal Live API and Thinking Mode for more dynamic use cases.
• Users are encouraged to explore these features via a Colab notebook, GitHub examples, and the Gemini API docs, and to join the Gemini API Developer forum to share feedback and ideas for improvements.

Happy building, and enjoy exploring what Gemini 2.0 can do for you!
Link to Article