Developers introduced the Realtime-Venus system for audiovisual communication

Team Realtime-Venus released a full-duplex system for audio-visual interaction with support for asynchronous task delegation, aimed at creating a natural dialogue between the user and AI in real-time.

The system is based on a model with 9 billion parameters, based on MiniCPM-o 4.5 or Qwen3-8B architectures, and its technical implementation supports proactive interaction, correct handling of interruptions during speech, and a long-term memory function for video content that works without the need for additional fine-tuning.

The system is designed to process streaming data, allowing the model to respond to visual and audio signals with minimal delays, and the use of the 9B architecture provides a balance between performance and computational resource requirements, allowing the solution to be deployed in various interaction scenarios.

The asynchronous task delegation function allows the system to perform background tasks without interrupting the current communication flow with the user, expanding the model's capabilities as a personal assistant that can maintain the context of long video sessions and quickly switch between tasks during the dialogue.

The release provides developers with a ready-made toolkit for integrating multimodal capabilities into third-party applications, and the source code and model weights are available in open access, allowing for local testing of the system and its adaptation to specific tasks of audio-visual stream processing.