Ant Group and Tsinghua University Open Full-Duplex Model Realtime-Venus

Venus (Ant Group) and Tsinghua University released Realtime-Venus, an open full-duplex interaction system with asynchronous background task delegation, publishing two 9B checkpoints under the Apache 2.0 license.

Compiled by Sergey KostenchukPublished 2026-09-26Updated 2026-09-26

2026-09-26 Coding NVIDIA

🤖 Ant Group and Tsinghua University Open Full-Duplex Model Realtime-Venus

Two 9B-parameter checkpoints based on MiniCPM-o 4.5 with a Qwen3-8B backbone have been released: Realtime-Venus-Omni (audio and video, autonomously deciding when to join the dialogue) and Realtime-Venus-Audio (voice). Background tasks via <delegate> queries are handled by the Realtime-Venus-Harness runtime without blocking the conversation, returning results to the same chat.

🌍 According to arXiv:2609.13814, Omni leads in 6 out of 8 video benchmarks (StreamingBench 70.2%), while Audio scores 78.0% on MMAU and 97% speech continuation during 'backchannels,' reacting to real interruptions in 75% of cases.

👤 Apache 2.0 weights and a beta Android APK allow local full-duplex deployment; researchers get an open object for studying interruptions and turn-taking. The online demo requires at least one NVIDIA A100.

Source 1: https://arxiv.org/abs/2609.13814 Source 2: https://github.com/inclusionAI/Realtime-Venus

Sources