llm-gemini plugin supports 3.7 Flash models and server-side tool experiments

Gemini models accelerate terminal integration

The release of the new llm-gemini plugin version 0.33 arrives with support for Google Gemini 3.7 Flash. In addition, it makes the lightweight 3.5-flash-lite model and embedding models available. The plugin serves as a bridge for developers who want to use Google's models directly from their terminal.

But the main draw doesn't lie merely in the list of supported models. Simon Willison has focused on compatibility with the new version of the LLM 0.32 tool.

Visible reasoning traces expose model bottlenecks

Compatibility with version 0.32 means two key changes. The first is the visibility of reasoning traces, the model's thought processes before generating the final answer. Thanks to this, developers can better fine-tune prompts and understand where the model gets stuck when processing more complex tasks.

The second shift is the support for running server-side tools. The model can now access other scripts right on the command line.

Executing code from CLI pushes risk boundaries

Integrating features like CodeExecution from the terminal environment pushes the logic of LLM tools. Instead of copying the code from the model and running it manually, you can directly command Gemini to generate and execute the necessary script, such as a Python equation from the demo.

The combination of faster 3.7 Flash models with these tools gives developers a very powerful set for rapid prototyping. However, it raises questions about local environment security.

Security friction will determine adoption

With his plugin, Simon Willison continues to build an ecosystem that bypasses web GUIs. For more experienced developers, it will be crucial to see whether integrating server-side functions into the terminal causes unacceptable risk.

Running generated code straight from the CLI without strict isolation introduces a vulnerability that production developers may not want to let near their setups.

Lilith's verdict

The terminal is no longer just a dumb input for a chatbot. It starts as a helper for developers and ends as a system that runs Python scripts by itself right on your machine.