Skip to main content

​
Base URL

​
Authentication

​
Endpoints

​
Models

Do not hardcode a catalog from this page. List IDs with GET /v1/models and pick one whose type_target is gemini. Image models such as Nano Banana use the same generateContent path; see Images.

​
Thinking and Thought Signatures

Gemini thinking models return an encrypted thoughtSignature on some response parts. With Gemini 3 models, the signature on a function call part must be sent back on the same part in the following requests of that turn; otherwise the model returns 400. For Gemini 2.5 models, returning it is optional. This is native model behavior, described in Google’s Thought signatures guide.
  • The official Google Gen AI SDKs handle signatures automatically when you append the full model response to the history. With REST, or when you rebuild history yourself, keep every part exactly as received.
  • A signature is validated by the upstream that issued it. PipeLLM may serve consecutive requests for the same model through different upstream channels. On the Responses converter route we observed a Gemini 3 signature replay rejected with 400 Thought signature is not valid after such a switch, while same-channel replays, including a native-route replay, succeeded. Handle this error in Gemini 3 tool loops.
  • Converter routes differ in how well they preserve signatures. See Protocol Conversion Overview.

​
Thinking configuration

Gemini 3 uses generationConfig.thinkingConfig.thinkingLevel; Gemini 2.5 uses thinkingBudget. Accepted levels vary by model. Native requests use the target model’s settings; converter mappings and restrictions are documented on each format page.

​
Next Steps

Generate Content

Endpoint reference and SDK examples

LangChain (Gemini)

Framework integration guide

Nano Banana

Image generation