Gemini 3.8 Live Reaches General Availability
Google released Gemini 3.8 Live for real-time voice applications on September 15, 2026. Learn about its low-latency design and evaluation steps.
Official Release Details
The source event date is September 15, 2026. On this date, Google declared Gemini 3.8 Live generally available. The release includes two new audio-to-audio models designed for real-time voice applications using the Live API. Gemini 3.8 Live is identified as the default option for most low-latency voice agent experiences and real-time dialogue without reasoning delays.
- Model name: Gemini 3.8 Live
- Provider: Google
- Input modalities: text, image
- Output modality: text
- Context window: 131,072 tokens
- Max output tokens: 65,536
Documented Features
The official changelog lists specific technical features for this release. The model supports interleaved reasoning, which allows for processing logic during conversation flow. It also includes default asynchronous function calling, enabling background execution of tasks. Additionally, the system provides full session client content updates to maintain state across interactions. These features are explicitly stated in the source material as part of the GA release.
Evaluation Procedure
Engineers can validate the model’s performance by designing controlled tests. First, test the latency response time during continuous voice dialogue sessions. Measure the delay between user input and model output to verify low-latency claims. Compare these metrics against baseline requirements for your specific application. Validate that asynchronous function calls execute correctly without blocking the main audio stream. Note that open weights are not confirmed for this model, so local installation capabilities remain unverified in the provided facts.
Primary sources
Frequently asked questions
What is the primary use case for Gemini 3.8 Live?
It is designed for real-time voice applications requiring low latency. It serves as the default option for voice agent experiences where reasoning delays must be minimized.
Are open weights available for this model?
The source indicates that open weights are not confirmed. Therefore, no statement can be made regarding local installation or self-hosting based on the provided facts.