Ant and Tsinghua unveil Realtime-Venus to run tasks while chatting

Ever asked a voice assistant to look something up, only to be met with dead silence while it thinks? Ant Group and Tsinghua University just unveiled Realtime-Venus to eliminate that awkward pause.

Chatting while working

Realtime-Venus is a real-time multimodal agent system designed to continuously see, hear, and talk while quietly handling complex workloads in the background.

At its core is a dual-loop architecture powered by two 9B models and a coordination layer called Harness. Realtime-Venus-Omni continuously processes incoming audio and video streams, while Realtime-Venus-Audio drives native full-duplex voice conversation.

When a task requires web search, heavy reasoning, or tool execution, Realtime-Venus-Harness delegates the job to the background asynchronously. Once complete, it routes the findings back into the live conversation, so users never have to pause talking just to wait for the agent to finish.

Why it matters

Decoupling the conversational front-end from heavy back-end tasks is a genuinely smart design choice. Traditional voice interfaces freeze up the moment an external tool is triggered, but Realtime-Venus behaves much more like a human assistant who can keep chatting while looking something up for you.

We think this architecture is a natural fit for smart glasses and in-car systems, where dead air and frozen interfaces completely ruin the experience. Running two 9B models alongside background agent routines will certainly test hardware power budgets and connectivity, but it sets a practical blueprint for ambient AI assistants.