XNews

Google Releases Gemini 3.8 Live and Extended Thinking

Google Releases Gemini 3.8 Live and Extended Thinking

Google launched Gemini 3.8 Live and 3.8 Live Extended Thinking, new state-of-the-art live audio models. 3.8 Live supports 97 languages with seamless switching, near real-time visual understanding, and background tool calls. Extended Thinking adds multi-step reasoning with narrated progress, targeting enterprise tasks.

Key insights

Gemini 3.8 Live Focuses on High-Concurrency, Low-Latency Deployment

The base Gemini 3.8 Live model is designed for cost-effective large-scale deployment. It supports high-concurrency, low-latency daily assistant and retrieval services. It can automatically detect and smoothly transition among 97 languages during conversation. The model also enables fluent multimodal real-time interaction combining vision and voice at competitive running costs.

Extended Thinking Version Adds Real-Time Reasoning with Narration

Gemini 3.8 Live Extended Thinking is aimed at enterprise-grade complex tasks. It strengthens multi-step logical reasoning and planning. The model achieves 'real-time thinking and speaking synchronization': when faced with a complex problem, it naturally says 'let me check...' as a transition, then narrates progress while executing multi-step tasks in the background. This creates a conversational experience similar to talking to a human on the phone.

Background Asynchronous Tool Calls Enable Concurrent Collaboration

Both models support background asynchronous tool calls. While calling external APIs, retrieving data, or booking reservations, the front-end voice chat continues uninterrupted. This enables true concurrent collaboration, allowing the model to perform actions without breaking the conversational flow.

Near Real-Time Visual Understanding Demonstrated in Multiple Scenarios

The models can watch a scene and chat simultaneously. Examples include answering questions about a new employee's onboarding environment, playing chess by looking at the board, and even looking at a sketch and using voice to generate React code. This capability is available in both Live and Extended Thinking versions.

Source posts