llm-gemini 0.33 Adds Gemini 3.7 Flash and Server-Side Tools
Simon Willison released llm-gemini 0.33 on August 13, adding Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and two embedding models to the LLM command-line plugin. The update also requires LLM 0.32 and moves Google Search, URL context, and code execution onto LLM's typed server-side tool interface while preserving Gemini thought signatures across tool-use turns.
Simon Willison released llm-gemini 0.33 on August 13, updating the plugin that connects the open-source LLM command-line tool to Google's Gemini API. The release adds direct model identifiers for Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite, plus support for the gemini-embedding-2 and gemini-embedding-001 embedding models.
Google's Gemini API changelog separately confirms that Gemini 3.7 Flash became generally available on August 13. That makes the plugin update immediately useful to developers who want to call the new model through LLM rather than build a separate Gemini client.
Tool use and reasoning become structured events
Version 0.33 requires LLM 0.32 or later and adopts its structured message and streaming interfaces. Reasoning output, tool calls, and tool results are emitted as typed stream events. The plugin also preserves Gemini thought signatures across tool-use turns and correctly replays stateless message histories, addressing details that matter in multi-step agent workflows.
Google Search, URL context, and code execution now use LLM's server-side tool interface. Developers can enable them with the GoogleSearch, URLContext, and CodeExecution tool names instead of relying on plugin-specific handling. The release notes also say native Gemini server-side tool activity is exposed as structured events and Gemini 3 models can combine those hosted tools with local functions.
Model and embedding cleanup
The update adds smaller-vector variants for both supported embedding models at 768 and 1,536 dimensions. Developers should treat a switch between gemini-embedding-2 and gemini-embedding-001 as a migration: the maintainer says their vector spaces are incompatible, so existing collections need to be re-embedded rather than mixed.
The registry also drops 35 unavailable model identifiers, including retired Gemini 1.5 and 2.0 entries and other experimental or preview models. For practitioners, the practical gain is a cleaner live model list and a more consistent event model for observing reasoning and tools in command-line experiments, evaluation scripts, and lightweight agent prototypes.
Key Points
- 1llm-gemini 0.33 adds Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and two supported embedding models to the LLM CLI plugin.
- 2The release requires LLM 0.32 and emits reasoning, server-side tool calls, and tool results as typed events while preserving Gemini thought signatures.
- 3The two embedding models use incompatible vector spaces, so developers switching between them must re-embed existing collections.
Scoring Rationale
The release gives LLM CLI users same-day access to Gemini 3.7 Flash and materially improves structured reasoning and tool-use support. Its impact is meaningful for developer workflows but narrower than the underlying model launch because it is an integration release.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
