Ant Group's Realtime-Venus 9B tops six of eight video tests

Ant Group's Venus Team put two open-weight 9B models on arXiv on September 12, 2026. The report claims Realtime-Venus-Omni, the audio-visual model, took 'the highest scores on six of eight video benchmarks, including StreamingBench (70.2%), OVO-Bench (64.7%), and Daily-Omni (81.3%).' The audio-only sibling, Realtime-Venus-Audio, 'leads the compared models on MMAU (78.0%), MMAU-Pro (63.2%), Llama Questions (83.8%), and Speech CMMLU (67.8%).'

The full-duplex numbers are where the paper picks a fight. It reports the system 'responds to 75% of user interruptions and achieves continuation rates of 97%, 88%, and 86% under backchannels, other-directed speech, and background speech, respectively, exceeding Gemini 3.1 Live and GPT-4o on all three continuation metrics.'

Architecturally, the authors describe 'a dual-loop runtime [that] coordinates live interaction with background reasoning and tool execution.' A third component, Realtime-Venus-Harness, runs tool calls asynchronously while the front-end model keeps the dialogue going, then folds results back into the conversation.

Every number is the authors' own; no independent replication has been posted yet. Weights and code are on Hugging Face at inclusionAI/Realtime-Venus.