LLM-Gemini Plugin 0.33 Adds Gemini 3.7 Flash Support

👁 118 reads

Simon Willison released llm-gemini 0.33 on 13 August 2026, updating his open-source plugin for Google’s Gemini model family with support for five newly available models and a framework upgrade that unlocks reasoning traces and server-side tool use.

The release covers Gemini 3.7 Flash - which Google also launched on the same day - along with gemini-3.6-flash, gemini-3.5-flash-lite, and two embedding models: gemini-embedding-2 and gemini-embedding-001. The plugin had not received an update for some time before this version, according to Willison’s notes.

What the upgrade to LLM 0.32 enables

Beyond model coverage, the more significant change may be compatibility with LLM 0.32, a framework update Willison released on 4 August 2026 that added reasoning traces, OpenAI Responses support, server-side tools, and smarter logging. With llm-gemini now matching that version, users can activate Gemini’s built-in CodeExecution tool directly from the command line. Willison’s example invocation reads:

llm -m gemini-3.7-flash -T CodeExecution ‘use python to calculate (factorial of 13) * 3’

The pattern lets the model run Python server-side and return the computed result rather than simply generating code text. The same -T flag mechanism applies to other server-side tools exposed by the API.

Reasoning trace visibility is the other practical gain: users can now inspect the chain-of-thought steps Gemini produces when thinking effort is enabled, which was not surfaced by the plugin before the LLM 0.32 compatibility work.

One behavioural change in Gemini 3.7 Flash itself is worth noting for users who relied on fine-grained thinking controls. The minimal thinking effort option available in Gemini 3.6 Flash has been removed in the 3.7 generation. The remaining levels are high, medium, and low. Willison tested all three by asking Gemini 3.7 Flash to draw pelicans riding bicycles - a long-running personal benchmark - and reported the high-effort result as “pretty great.”

A correction posted on 14 August 2026 clarified an earlier claim: Willison had initially reported that the SVG output from Gemini 3.7 Flash rendered incorrectly in Chrome and Firefox, attributing the problem to invalid SVG from the model. He later established the rendering glitch was caused by a bug in his own rendering tool, not in the model’s output, and updated the post accordingly.

The plugin is available through the standard LLM plugin installation mechanism and requires LLM 0.32 or later for the reasoning trace and server-side tool features to function.

// read moreGoogle releases Gemini 3.7 Flash at half the price of predecessorNewsLLM 0.33 Adds Repeatable Templates, OpenAI Library 3.x SupportAI ToolsGemini Omni 1.2 Flash Pricing: Cost per Second and 360p to 4KAI Tools

Kinetyra