Realtime-Venus is a full-duplex interaction system (full-duplex meaning both sides can speak and listen at the same time) that handles the kind of natural, interruption-friendly back-and-forth that happens between humans and machines in real time, while also running actual work (tool execution) asynchronously in the background. It comes as two 9B-parameter (9 billion parameter) model variants: Realtime-Venus-Omni for speech and video, and Realtime-Venus-Audio for speech only. Each session uses one of the two, and Omni on its own handles both speech and video. According to the abstract, it scores best on six of eight video benchmarks and responds to 75% of user interruptions — a quantitative demonstration of its real-time conversational control.
Anyone who has talked to a voice assistant knows the frustration. It waits until you finish speaking, only then starts processing, and answers a good while later. It's like a walkie-talkie: "one person talks, the other only listens." This is called half-duplex.
But real human conversation doesn't work that way. While the other person is still talking, we drop in backchannels like "oh, I see" or "mm-hmm," and when something important comes up we cut in and steer the direction. We naturally filter out what someone nearby is saying or the background noise. When listening and speaking happen at once like this, that's full-duplex.
On top of this, two bigger challenges pile on.
First, there are situations where you also have to watch video. A voice conversation works off sound and linguistic cues alone, but to converse while seeing the world through a camera you have to ground the dialogue in "the shifting visual context unfolding right in front of you." The frame keeps flowing, and you can't afford to lose that flow.
Second, you have to do actual work while talking. A human assistant keeps the conversation going while searching for something or booking a reservation at the same time. But when an AI dives into heavy reasoning or tool execution, the conversation just stops dead. The abstract sets out to solve exactly these three things — continuous perception, real-time conversational control, and concurrent background work — within a single system.
Realtime-Venus's design can be unpacked with a few analogies.
1) One schedule (a shared causal timeline)
Each model in this system is not a mere speech recognizer but a "complete conversational frontend." It ties together continuous perception, conversational control, and native speech generation (native speech generation meaning the model produces the audio itself rather than handing text off to a separate speech synthesizer) into one.
What makes this bundling possible is a shared causal timeline. Think of a clerk recording meeting minutes in chronological order. What the user said, how the model responded, and even the delegation events for handed-off tasks are all recorded in order along a single time axis. "Causal" means it judges based only on the past and present without peeking at the future — a property essential for real-time streaming.
2) Two loops (a dual-loop runtime)
The key mechanism is a dual-loop runtime. Think of a restaurant's floor staff and kitchen and it gets easy.
When a guest places an order, the server says "sure, I'll get that going" and keeps carrying on other conversation. When the kitchen (Harness) finishes cooking, it hands the result back to the floor, and it gets woven naturally into the ongoing conversation. This is asynchronous delegation. It's exactly why the conversation doesn't stall because of a task.
3) A shared post-training recipe
The two models (Omni, Audio) are trained separately but share the same post-training approach. According to the abstract, three ingredients are mixed in: offline understanding (building up baseline comprehension from non-real-time material), proactive full-duplex trajectories (learning the flow of real-time conversation — when to cut in and when to hold back), and delegation workflows (learning when to hand work off to the background).
Here are the figures the abstract presents. All values below are taken only from what the abstract explicitly states.
Video understanding (Realtime-Venus-Omni) — best score on six of eight video benchmarks among the online models evaluated.
| Benchmark | Score |
|---|---|
| StreamingBench | 70.2% |
| OVO-Bench | 64.7% |
| Daily-Omni | 81.3% |
Audio understanding and spoken QA (Realtime-Venus-Audio) — evaluated on eight benchmarks, with sole best score on four of them and a tied best on one (it trails other models on MMAR, MMSU, and Speech TriviaQA).
| Benchmark | Score |
|---|---|
| MMAU | 78.0% |
| MMAU-Pro | 63.2% |
| Llama Questions | 83.8% |
| Speech CMMLU | 67.8% |
| VoiceBench AlpacaEval | 4.81 (tied for best score) |
Full-duplex interaction (Full-Duplex-Bench v1.5) — this is the paper's central differentiator.
The abstract states that it surpasses Gemini 3.1 Live and GPT-4o on all three of these continuation rate metrics. That edge, though, is confined to the non-interruption situations — the discernment to hold the conversation steady without cutting in when there's no need to. On reacting to genuine interruptions it does not lead: the paper's detailed results (§7.5) place its 75% interruption response rate below both Gemini 3.1 Live (77%) and GPT-4o (78%), and the paper itself notes that the advantage lies in the continuation rates. The significance is that it breaks away from walkie-talkie-style AI and takes a step toward the rhythm of human conversation.
Here's an honest accounting of what the abstract alone can't confirm.
ONTLE is a service that provides AI PMs on a deployment model. The structure of this paper overlaps strikingly with the direction we're aiming for.
First, the dual loop is essentially the basic skeleton of a "PM who actually works." A good PM runs meetings (foreground conversation) while simultaneously running research and document work (background execution) behind the scenes. An agent that stops mid-conversation and goes silent with "one moment, I'm working on it" fails as an assistant. The asynchronous delegation pattern — separating foreground responses from background tool execution and plugging finished results naturally back into the conversation — is worth treating as a reference standard for agent product design.
Second, "the discernment not to interrupt" is what separates good conversational agents from bad ones. The 86% continuation rate on background speech and 88% on other-directed speech quantify the ability not to ruin a conversation by reacting to every sound nearby. For an AI PM sitting in on a meeting, this ability to distinguish what's directed at it from what others are saying among themselves will make or break real-world satisfaction.
Third, the shared timeline offers a hint for agent logging and RAG design. The idea of recording user input, model output, and delegation events on a single causal time axis is a practical blueprint for how to accumulate an agent's action history and retrieval (RAG) context consistently. When what was known, when, and when something was delegated all remains in a single thread, after-the-fact auditing and reproduction become easy.
That said, in line with ONTLE's positioning, this technology should be read not as a tool to replace people but as hands and feet that lighten repetitive work. The better it gets at listening, watching, and delegating in real time, the more room people have to focus on judgment — and that is the real product value.
Existing voice assistants are mostly half-duplex (walkie-talkie style), so they respond only after the user finishes speaking. Full-duplex has listening and speaking happen at the same time, so it can react in real time to backchannels or interruptions while the user is still talking. Per the abstract, this system responds to 75% of user interruptions.
Realtime-Venus-Omni is for speech-and-video interaction and Realtime-Venus-Audio is speech-only, each trained separately at 9B size. According to the paper, each session uses one of the two (you don't run both at once), and Omni handles speech and video together. The two variants share the same post-training recipe (offline understanding + full-duplex trajectories + delegation workflows).
It's a way of handling heavy reasoning or tool execution separately in the background (Realtime-Venus-Harness) without stopping the conversation, then weaving the result back into the ongoing dialogue. It's like a restaurant server passing an order to the kitchen, continuing to chat with the guest, and bringing the dish out naturally once it's ready.
The abstract states that it surpasses both Gemini 3.1 Live and GPT-4o on the three conversation continuation rate metrics of Full-Duplex-Bench v1.5 (backchannel, other-directed speech, background speech). However, this is a comparison limited to specific full-duplex metrics and does not mean it is generally superior across all areas. On the interruption response rate, in fact, the paper's detailed results place it behind both commercial models.