inclusionAI has released the Realtime-Venus system on Hugging Face, featuring two checkpoints: Realtime-Venus-Omni and Realtime-Venus-Audio. The Omni checkpoint is a 9B audio-visual interaction model adapted from MiniCPM-o 4.5 that continuously watches and listens to decide when to respond.

  • Native full-duplex conversation allows the model to perceive while speaking, distinguishing backchannels and interruptions.
  • Omni-Proactive interaction initiates responses based on temporally aligned video and audio events without waiting for a user prompt.
  • Delegation emits in-stream requests on a shared causal timeline to handle external tasks without blocking the conversation.
  • Training-free long-video memory archives visually informative moments and retrieves relevant evidence without additional training.
  • The system generates both text and speech output using bundled Token2wav resources.

The release provides tools for proactive, real-time audio-visual interaction with semantic interruption handling and long-term memory capabilities.