Glonce

llm-gemini 0.33 adds Gemini 3.7 Flash, reasoning traces

toolsThe Glonce editorsSimon Willison🔥47Glonce rating47Impact35Novelty28Relevance87Surprise12Players52Credibility88
llm-gemini 0.33 adds Gemini 3.7 Flash, reasoning traces

Simon Willison has released version 0.33 of llm-gemini, the plugin that connects Google's Gemini models to his command-line tool, llm. The update adds support for Gemini 3.7 Flash, which came out the same day as this release, along with gemini-3.6-flash, gemini-3.5-flash-lite, and two embedding models, gemini-embedding-2 and gemini-embedding-001.

The release also upgrades llm-gemini for compatibility with LLM 0.32, the underlying tool's own recent update. That brings two new capabilities: reasoning traces become visible, and server-side tools can be enabled. Willison demonstrates the tool-calling feature with a single command that turns on Gemini's CodeExecution tool and asks it to compute a factorial: llm -m gemini-3.7-flash -T CodeExecution 'use python to calculate (factorial of 13) * 3'. The Python calculation then runs as a server-side tool call rather than locally on the user's machine.

To show off the new model, Willison had Gemini 3.7 Flash draw pelicans riding bicycles at three thinking-effort settings: high, medium, and low. The 'minimal' effort option that existed in Gemini 3.6 Flash has been removed in Gemini 3.7 Flash. Only the high-effort image is shown in the post, which Willison calls 'pretty great.'

That image comes with an unrelated rendering catch. Willison says the pelican picture shown was rendered in Safari, and that both Firefox and Chrome render the same output differently, because Safari is more tolerant of empty SVG filter elements than the other two browsers are. In Firefox and Chrome, the bicycle still shows up, but the pelican itself is missing from the image entirely.

Key facts

  • llm-gemini 0.33 adds support for Gemini 3.7 Flash, released the same day, plus gemini-3.6-flash, gemini-3.5-flash-lite, and two embedding models, gemini-embedding-2 and gemini-embedding-001.
  • The plugin is upgraded for compatibility with LLM 0.32, adding visibility into reasoning traces and letting users enable server-side tools, shown with a command that runs Gemini's CodeExecution tool to compute a factorial in Python.
  • The 'minimal' thinking-effort setting available in Gemini 3.6 Flash has been removed in Gemini 3.7 Flash, leaving high, medium, and low.
  • Willison shows a high-thinking-effort SVG of a pelican riding a bicycle drawn by Gemini 3.7 Flash, calling it 'pretty great'; the medium and low-effort versions are not shown.
  • The showcased image was rendered in Safari; in Firefox and Chrome the same SVG shows the bicycle but not the pelican, because Safari tolerates empty SVG filter elements that the other two browsers do not.

Why it matters

llm-gemini is the plugin that connects Google's Gemini models to Willison's llm command-line tool, so version 0.33 means Gemini 3.7 Flash, gemini-3.6-flash, gemini-3.5-flash-lite, and the two new embedding models are usable from the terminal the same day Gemini 3.7 Flash itself shipped, without a separate SDK update. The LLM 0.32 compatibility upgrade adds two capabilities beyond model support: reasoning traces become visible, and server-side tools, demonstrated here with Gemini's CodeExecution tool, can be triggered with a single command instead of custom integration code.

Who it affects

Developers and other users of Willison's llm command-line tool who want Gemini 3.7 Flash, gemini-3.6-flash, gemini-3.5-flash-lite, or the two new embedding models, gemini-embedding-2 and gemini-embedding-001, available without switching tools or waiting on a separate SDK; and anyone building terminal-based workflows who wants to inspect Gemini's reasoning traces or trigger its server-side tools, such as CodeExecution, directly from a command line.

How to use it

Upgrading to llm-gemini 0.33 unlocks the new models by name: gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-embedding-2, and gemini-embedding-001. Gemini 3.7 Flash supports three thinking-effort levels, high, medium, and low; the 'minimal' option from Gemini 3.6 Flash is gone. To call a server-side tool, add a -T flag with the tool's name; Willison's own example enables CodeExecution and asks it to compute a factorial: llm -m gemini-3.7-flash -T CodeExecution 'use python to calculate (factorial of 13) * 3', which runs the Python as a server-side call rather than locally. The post gives no pricing, benchmark, or accuracy figures for any of the five new models.

How solid is it

This is a first-party release note written in the first person about llm-gemini's own new version, not a secondhand report, so the list of what changed in 0.33 is about as directly sourced as it gets. It reads as an informal blog post rather than a structured changelog: it gives no exact date for the release itself, only that it had 'been a while' since the last one and that Gemini 3.7 Flash shipped 'today'; it shows only the high-thinking-effort pelican drawing, while the medium and low-effort versions it mentions testing are neither shown nor described; and no benchmark, pricing, or accuracy figures back any of the five newly supported models.

Risks and caveats

No calendar date is attached to the llm-gemini 0.33 release itself, and no reason is given for why the 'minimal' thinking-effort option was dropped between Gemini 3.6 Flash and Gemini 3.7 Flash. None of the five newly supported models, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-embedding-2, and gemini-embedding-001, come with pricing, benchmark, or accuracy figures in the post. The rendering mismatch it describes, the same SVG showing a full pelican on a bicycle in Safari but only a bicycle in Firefox and Chrome, is a reminder that a single-browser screenshot of model-generated SVG output does not necessarily represent how it renders elsewhere.

“Both Firefox and Chrome render it differently, due to Safari being more tolerant of empty SVG elements than those other two browsers.”

— Simon Willison