Gemini 3.x, Google GenAI, multimodality, and grounding

The current stable Gemini family includes gemini-3.6-flash for balanced agentic/multimodal work and gemini-3.5-flash-lite for low-cost high-volume automation. Preview IDs are not stable production dependencies.

Spring AI integration

Spring AI 2.0 uses the Google GenAI SDK integration:

<dependency>
  <groupId>org.springframework.ai</groupId>
  <artifactId>spring-ai-starter-model-google-genai</artifactId>
</dependency>

Use provider options for thinking, multimodal inputs, native tools, grounding, and features not represented by common ChatOptions.

Tools and structured output

Gemini can use managed Google Search, Maps, URL context, File Search, and code execution, plus custom function calls. Function calling requests an intermediate action; structured output constrains the final answer. Preserve all required model response parts and tool IDs across a manual loop.

Embedding 2

gemini-embedding-2 is a GA multimodal embedding model for text, images, video, audio, and PDF. It accepts up to 8,192 input tokens and produces 128–3072 dimensions; 768, 1536, and 3072 are recommended. Truncated vectors are automatically normalized.

Tuning reality

Gemini API model tuning is not currently supported: the last tuning model was shut down in May 2025. Use prompts, RAG, tools, or another approved training path; do not design around a dead endpoint.

Lifecycle discipline

Read the Gemini changelog and deprecation page automatically in release management. Model aliases, parameters, and preview capabilities can change.

Feynman check

Gemini is a multimodal engine with Google tools. Explain why a stable model ID, an embedding model, and a managed research agent are three different products.