Realtime-Venus runs tools in the background without pausing the conversation

single source· 1 articles · confidence: medium · first seen 2026-09-11 20:00 UTC

What this means for you

Nothing to build on yet — no weights, code or API appear in the abstract. If you are comparing full-duplex voice stacks, the figure to check first is interruption handling: it answers 75% of interruptions, measured against two named commercial systems whose evaluation setup is not described.

A preprint posted on 11 September describes Realtime-Venus, a full-duplex interaction system (both sides can speak and listen at once) built from two separately trained 9B models — one for audio and video, one for speech alone. A second loop, Realtime-Venus-Harness, runs tools in the background and feeds results back into the live conversation. The authors report the video model scoring 70.2% on StreamingBench and 81.3% on Daily-Omni, and the audio model 78.0% on MMAU. The abstract gives no evaluation dates and does not say whether weights, code or an API will be released.

Key facts

  • ·Realtime-Venus is built from two separately trained 9B models: Realtime-Venus-Omni for audio-visual interaction and Realtime-Venus-Audio for spoken interaction. source
  • ·The authors report Realtime-Venus-Omni as highest among the evaluated online models on six of eight video benchmarks, including StreamingBench (70.2%), OVO-Bench (64.7%) and Daily-Omni (81.3%); no evaluation date is given. source
  • ·Realtime-Venus-Audio is reported at 78.0% on MMAU, 63.2% on MMAU-Pro, 83.8% on Llama Questions and 67.8% on Speech CMMLU, and matches a VoiceBench AlpacaEval score of 4.81. source
  • ·On Full-Duplex-Bench v1.5 the audio model responds to 75% of user interruptions and continues on 97%, 88% and 86% of backchannel, other-directed-speech and background-speech turns, which the authors say exceeds Gemini 3.1 Live and GPT-4o on all three continuation metrics. source
  • ·The system uses a dual-loop runtime in which foreground interaction continues while Realtime-Venus-Harness executes tasks asynchronously and returns results into the ongoing dialogue. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire