Code execution offers Gemini fashions entry to a Python sandbox, permitting the fashions to run code and study from the outcomes. With code execution enabled, Gemini fashions can carry out calculations, analyze advanced knowledge units, and create visualizations on the fly – finally offering higher solutions to person queries. It’s now typically obtainable in each Google AI Studio and the Gemini API through the Gemini 2.0 fashions.
Code execution as a software
You possibly can allow code execution through toggle in Google AI Studio’s “Instruments” panel, or from the Gemini API as a instruments variable (proven beneath).
from google import genai
from google.genai import sorts
shopper = genai.Shopper(api_key="GEMINI_API_KEY")
response = shopper.fashions.generate_content(
mannequin='gemini-2.0-flash',
contents="""
What's the sum of the primary 50 prime numbers?
Generate and run code for the calculation.
""",
config=sorts.GenerateContentConfig(
instruments=[types.Tool(
code_execution=types.ToolCodeExecution
)]
)
)
When code execution is added as a software, the mannequin can entry the code execution sandbox to run code for as much as 30 seconds at a time, and as much as 5 occasions with out re-prompting. The code execution setting consists of libraries like Numpy, Pandas, and Matplotlib (for graph rendering). The total checklist of accessible libraries will be present in our API docs and we need to increase the set of supported libraries quickly.
File IO and Graph Output
With Gemini 2.0, we’ve up to date our code execution software to permit file enter into the code execution sandbox, in addition to graphs & chart output utilizing Matplotlib. These updates result in a broader set of use circumstances for code execution. With these updates now you can:
- Carry out logical evaluation to grasp advanced user-uploaded recordsdata.
- Visualize knowledge with charts and graphs supported by Matplotlib.
- [Experimental] Unlock code execution’s real-time capabilities with the Multimodal Stay API.
- [Experimental] Mix code execution with instruments like Grounding with Google Search.
- [Experimental] Leverage code execution from Gemini 2.0 Pondering Mode.
Let us take a look at two sensible examples of code execution in motion:
Actual-time knowledge evaluation and visualization with Gemini fashions
By combining the Multimodal Live API and code execution, this demo demonstrates a dwell dialog with the Gemini mannequin, using each voice enter and code execution. We ask Gemini fashions to supply a listing of Tom Cruise films, rank them by runtime, after which create a bar chart visualizing this knowledge utilizing Matplotlib. Gemini fashions generate the Python code to carry out these duties and replace the chart based mostly on our additional requests (altering bar colours).
(Proper-click and open video in new window to listen to audio)
Combining Pondering mannequin and code execution to unravel advanced issues
This demo makes use of the Gemini 2.0 Flash Pondering Experimental mannequin and code execution to unravel a traditional optimization problem. We ask Gemini to seek out the shortest route for a salesman visiting 5 Spanish cities and returning to their start line. Gemini generates Python code, debugs it iteratively (correcting an preliminary library error), calculates distances, and finally visualizes the optimum route on a Matplotlib graph.
(Proper-click and open video in new window to listen to audio)
Get began with Gemini 2.0 Code Execution
Need to attempt your self? Take a look at our Colab notebook to get hands-on with code execution or the GitHub For additional examples and demos please go to the Gemini API docs, the place you may discover eventualities of code execution getting used and extra.
We welcome you to affix the Gemini API Developer forum to share your use circumstances and suggestions on how we will enhance code execution for you. Within the quick time period, we’re exploring expanded library help, supporting different enter modalities like PDFs, and supporting multi-tool use. Glad constructing!
