For the complete documentation index, see llms.txt. This page is also available as Markdown.

Gemini 3.8 Live

This documentation is valid for the following list of our models:

  • google/gemini-3.8-live

  • gemini-3.8-live

Google's Live API model for real-time voice and video conversations: it takes streamed speech, text and camera or screen frames, and answers with speech or text, with function calling. Sessions speak Google's Gemini Live protocol, so the official Google GenAI SDKs work with only the base URL changed.

Setup your API Key

If you don’t have an API key for the AI/ML API yet, feel free to use our Quickstart guide.

API Schema

The model is served over a WebSocket in the Gemini Live protocol. Client messages, billing, limits and errors are described in Realtime Sessions.

Connection URL (API key)

wss://api.aimlapi.com/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent (v1alpha also works), with the key in x-goog-api-key or Authorization: Bearer

Connection URL (client secret)

wss://api.aimlapi.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContentConstrained (v1beta also works), with the secret as access_token or in Authorization: Token (client secrets)

SDK base URL

https://api.aimlapi.com

JavaScript SDK

@google/genai sends a plain API key in the URL, which is refused (400 credential_in_query): pass a client secret as its API key with apiVersion: "v1alpha", as in the example below

Session options (setup, the first message, sent once within 10 seconds of connecting; model is gemini-3.8-live, google/gemini-3.8-live, models/gemini-3.8-live or models/google/gemini-3.8-live):

Option
Values

generationConfig.responseModalities

AUDIO (default) and/or TEXT

generationConfig.maxOutputTokens

Kept between 1,024 and 4,096 (default 4,096); thinking counts inside it

generationConfig.temperature, topP, topK, presencePenalty, frequencyPenalty, seed

As in the Gemini API; candidateCount 1 only

generationConfig.speechConfig

voiceConfig.prebuiltVoiceConfig.voiceName (a prebuilt voice such as Puck or Kore), languageCode

generationConfig.mediaResolution

MEDIA_RESOLUTION_LOW, _MEDIUM, _HIGH

systemInstruction

Up to 8 text parts of 16,384 characters

tools

functionDeclarations only, up to 32

realtimeInputConfig

automaticActivityDetection, activityHandling, turnCoverage

outputAudioTranscription

Transcript of the model's speech

historyConfig.initialHistoryInClientContent

true or false

Code Example

Response

The example sends one typed turn and collects the spoken answer (raw PCM16 audio) until the turn is complete. The Python version prints the token usage Google reports and the size of the received audio. Values in angle brackets are placeholders.

Last updated

Was this helpful?