0.33

• Updated for LLM 0.32's structured message and streaming APIs. Reasoning, tool calls and tool results are now emitted as typed stream events, Gemini thought signatures are preserved and replayed across tool-use turns and stateless messages= histories are replayed correctly. This release requires LLM 0.32 or later. #132
• New model gemini-3.7-flash for Gemini 3.7 Flash. #145
• New models gemini-3.6-flash and gemini-3.5-flash-lite for Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. #139
• Added the gemini-embedding-2 and gemini-embedding-001 embedding models, each with -768 and -1536 variants for smaller vectors. The deprecated text-embedding-004 and gemini-embedding-exp-03-07 models have been removed. The two supported models use incompatible vector spaces, so existing collections must be re-embedded when switching between them. Thanks, John Blomberg. #138
• Google Search, URL context and code execution now use LLM's server-side tool interface and can be enabled using -T GoogleSearch, -T URLContext and -T CodeExecution. Native Gemini server-side tool calls and results are exposed as structured events, and Gemini 3 models can combine server-side tools with local function tools. #141
• Google Search grounding now retains Gemini's raw groundingMetadata without altering the model's response text. Search suggestions are emitted as display-only events—shown on standard error by the CLI—so they are excluded from response text, logs and subsequent conversation turns. Thanks, Andrew Hoddinott. #40
• Removed 35 unavailable Gemini models, including retired Gemini 1.5 and 2.0, Gemma 3, experimental and preview models. The registry now reflects model IDs verified to accept live generateContent requests. #142