llm-gemini 0.33 adds Gemini 3.7 Flash support, reasoning traces

llm-gemini, a plugin for the LLM command-line tool, released version 0.33 after a period without updates. The plugin now supports Gemini 3.7 Flash, Google's latest iteration of the Flash reasoning model, alongside three other Gemini models: gemini-3.6-flash, gemini-3.5-flash-lite, and two embedding models, gemini-embedding-2 and gemini-embedding-001.

The update bumps compatibility to LLM 0.32, which enables two powerful features for developers. First, reasoning traces are now visible—developers can see the intermediate steps and logic the model uses when solving a problem. Second, server-side tools can be enabled, allowing code execution (such as Python) to run on the server within the model's context. An example command shows how to calculate (factorial of 13) × 3 in Python: `llm -m gemini-3.7-flash -T CodeExecution 'use python to calculate (factorial of 13) * 3'`.

One notable change in 3.7 Flash is the removal of the "minimal" thinking-effort option that was present in 3.6 Flash; now only high and medium thinking efforts are available. To test the new model's capabilities, the author prompted Gemini 3.7 Flash to render pelicans riding bicycles at different thinking-effort levels. The high-effort render produced detailed results, but the author noted a quirk: when rendered in Safari, the SVG image displays correctly, but in Firefox and Chrome the pelican disappears entirely, though the bicycle remains visible. This discrepancy stems from Safari's greater tolerance for empty SVG filter elements compared to the other two browsers.