A generated title card for the article 'Avatar 6 vs Realtime Venus', with a kicker reading 'AVATAR MODEL COMPARISON', the large title 'Avatar 6 vs Realtime Venus', the subtitle 'A renderer and a listener, compared by mistake', and three pill cards reading 'video, up to 2K', 'speech, Apache-2.0' and 'different jobs'.
Engineering & Research

Avatar 6 vs Realtime Venus: A Renderer and a Listener, Compared by Mistake

Author

Elias Hawthorne

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Put Avatar 6 and Realtime Venus in the same sentence and the first thing that needs saying is that one of them cannot show you a face. Realtime Venus is Ant Group's full-duplex interaction model — a 9B system that listens while it speaks, understands sound in the context of what it can see, and delegates work in the background. Its outputs are speech and decisions. It does not generate video, does not render a presenter, and has no concept of a reference image. Avatar 6, which HeyGen released on 17 November 2026, does exactly one thing: it animates a person from a photo or a short clip, driven by audio you supply, and returns video up to 2K.

So the comparison is a category error, and the useful version of it is a design question. If you are building something where a human appears on screen and holds up their end of a conversation, you need both kinds of system, and the decision that matters is which side of the stack you own. That is where these two diverge sharply — one you can download, the other you can only rent.

Two portfolios of evidence, both honest about their limits

Realtime Venus arrives the way open research arrives: an arXiv report (2609.13814), Apache-2.0 weights on Hugging Face under inclusionAI/Realtime-Venus, two 9B checkpoints, and a project page with demos. There was no launch post, no API, no price and no benchmark that an outside party has reproduced. Every number attached to it comes from the technical report the authors wrote about their own model.

Avatar 6 comes with the opposite apparatus: a product page, two identifiers, a REST endpoint, a free tier, and a technical report whose preference studies were all run by HeyGen on HeyGen's raters. It does have one thing Venus lacks — an outside measurement. New Measure's blind arena, published 18 November, put it first on both of its boards at 1,197 and 1,237 Elo, over clips no longer than 35 seconds.

Read both headline figures as claims rather than findings. HeyGen's 84.4% is a 42.2% side-by-side win rate rescaled against a 50% baseline; its own single-clip study puts Avatar 6 at 65.1% judged real against 78.9% for unedited footage. Venus's 97% backchannel continuation comes from Full-Duplex-Bench v1.5, a protocol its authors chose. Neither number is dishonest, and neither has been replicated by a party with nothing to sell — though Avatar 6 has at least been voted on by strangers, for half a minute at a time.

A generated two-column scoreboard titled 'Avatar 6 vs Realtime Venus — the scoreboard'. Left column 'Avatar 6': Backbone undisclosed; Input reference media plus audio; Output video up to 2880x1620; Turn-taking none, it performs a script; Duration 25-28 min demonstrated; Access and licence HeyGen credits. Right column 'Realtime Venus': Backbone 9B omni: SigLIP2, Whisper-Medium, Qwen3-8B; Input continuous audio and video streams; Output speech and text, no pixels; Turn-taking full duplex, 97% backchannel; Duration a running session, 40,960-token context; Access and licence Apache-2.0 weights. The footer reads 'Venus scores self-reported in arXiv 2609.13814. Avatar 6 per HeyGen plus New Measure.'

What each system is made of

• Backbone — Avatar 6: undisclosed; HeyGen describes it as its smartest avatar model and documents no architecture. Realtime Venus: a 9B omni model with SigLIP2 vision, Whisper-Medium audio and a Qwen3-8B language backbone, adapted from MiniCPM-o 4.5 per the model card.

• Modality in — Avatar 6: one reference image or one reference video plus speech audio. Realtime Venus: continuous audio and video streams, plus delegated task context.

• Modality out — Avatar 6: video with synced speech, up to 2880×1620 through the Model API. Realtime Venus: speech, text, and control signals; no pixels.

• Duration — Avatar 6: as long as the audio, with 25-to-28-minute continuous renders demonstrated. Realtime Venus: a running session, with a 40,960-token context and training-free long-video memory that stores and retrieves selected frames.

• Turn-taking — Avatar 6: none. It performs a script; there is nobody to interrupt. Realtime Venus: full duplex, with measured interruption response and the ability to keep speaking through a backchannel.

• Access — Avatar 6: REST API, consent-gated, billed in HeyGen credits. Realtime Venus: weights on Hugging Face and ModelScope, no hosted endpoint, no price.

• Licence — Avatar 6: commercial terms under HeyGen's account. Realtime Venus: Apache-2.0.

That last row is the one that reorganises the comparison. Avatar 6 is a service. Realtime Venus is a model. Everything else follows from it.

A screenshot of HeyGen's research page for 'Avatar 6: Talking Avatars That Are More Like You', dated 17 November 2026 and marked a 17-minute read, with the abstract stating that Avatar 6 reproduces how a person looks, speaks and moves from a single photo, up to 2K, with no length limit.

The procurement difference is the real difference

If you want Realtime Venus in production, you download it, provision GPUs, and own the serving problem. Two 9B checkpoints, BF16, an ONNX export alongside the safetensors, an Android APK for on-device experiments, and a separate harness repository for the delegation loop. Nobody will send you a bill and nobody will deprecate your endpoint, but nobody will fix your latency either. For a research team that is a feature. For a product team it is a quarter of work before the first useful output.

If you want Avatar 6 in production, you sign up, get a key, and call POST /v3/models/avatars with a reference image and an audio URL. You get a video_id back in a 202, you poll, and a signed URL appears. The consent flow is heavier than most vendors impose — a realistic human reference video needs Enterprise access and an on-camera statement containing a session-specific one-time code — but the integration is an afternoon, not a quarter.

What you cannot do with Avatar 6 is price it per unit. HeyGen publishes $0.02 per second for HeyGen Video 1 and nothing equivalent for Avatar 6; the public pricing page sells Creator at $29/month for 600 credits and Pro at $49/month for 1,000 credits, with no credit-to-second conversion published. Realtime Venus has the opposite problem: the marginal cost of a call is your electricity bill, and the fixed cost is a GPU you have to justify to someone.

Where the two would actually meet

Picture the product both of these are components of: a person asks a question by voice, something decides what to say, and a face says it back with the right expression and a gesture on the important word. Realtime Venus covers the first two steps — it hears you, decides when to speak, and can keep listening while it talks. Avatar 6 covers the last one, and does it better than anything else currently purchasable, according to the only independent board that has looked.

The trouble is the seam. Avatar 6 wants a complete audio file — the video runs the length of the audio you hand it, and there is no streaming mode, no partial-audio input, no way to append. Realtime Venus produces speech incrementally, as the conversation moves. Joining them means buffering Venus's output into a finished utterance, shipping it to HeyGen, waiting out a render that the independent arena measured at a typical 175.6 seconds, and playing back a video that was generated after the moment had passed. That is not a conversation. It is a mail-order conversation.

For the real-time version you need a renderer that generates video as speech arrives, and that is a different product. Meta's Muse Realtime Avatar is the closest shipping example; Tavus Griffin, in research preview, is another. Avatar 6 is not that, and HeyGen has never claimed it was — its latency budget is a batch budget, and the arena's data bears that out: one in ten HeyGen personal-avatar renders took more than nine and a half minutes.

What Venus measures that nothing else here does

The interesting numbers on the Venus report are not quality scores. They are behavioural ones, and they describe a problem Avatar 6 does not have because Avatar 6 is not listening.

Full-Duplex-Bench v1.5 puts continuation at 97% for user backchannels, 88% for speech directed at others, and 86% for background speech, with a separate 75% response rate to genuine interruptions. The point of those four numbers together is that the model distinguishes "mm-hm" from "wait, stop" — it keeps going through the first and yields to the second. Understanding results run 70.2% on StreamingBench, 64.7% on OVO-Bench, 81.3% on Daily-Omni and 78.0% on MMAU audio. Delegation accuracy on the team's in-house benchmark is 75.93% for the Omni checkpoint and 68.89% for the Audio one, which measures whether the model correctly decides to hand a task off, not whether the task succeeds.

Those are useful engineering targets for anyone building a conversational agent, and they are unreproduced. The honest reading is that Venus is a credible open baseline for full-duplex behaviour, published with its weights so that the claim can be checked, and that nobody outside Ant Group and Tsinghua has published a check.

Where OrcaRouter sits

Neither of these is ours to sell. Realtime Venus is weights and a paper with nothing listening behind them; Avatar 6 is HeyGen's renderer, gated by HeyGen's consent flow and paid for in HeyGen credits. A routing pitch that blurred either line would be a lie, and it is the kind that surfaces in your first sprint rather than in a sales deck.

What we do serve is the delegated-reasoning leg that both architectures depend on. Venus's own design routes background tasks to a multimodal model, a general model, or a registered skill; Avatar 6's script has to be written by something. That layer is where a single key across 200+ models, at provider list price with no markup, and with automatic failover, actually removes work — because it is the only part of either pipeline where you have a choice of supplier.

A screenshot of the OrcaRouter Models page, showing the filter sidebar and the header reading 226 models, with the tabs All 226, Cyber 11, Text 190 and Image 12.

The reading that survives contact with a roadmap

If your problem is "we need a presenter on screen saying our words, at scale, this quarter," Avatar 6 is the answer and Realtime Venus is not in the conversation. Accept the three unknowns — no published per-second price, no independent test beyond 35 seconds, and a typical render time measured in minutes — and pilot it.

If your problem is "we need something that can hold a spoken conversation and decide when to speak," Realtime Venus is a serious open baseline and Avatar 6 is irrelevant, because it has no ears. Take the Apache-2.0 weights, plan for the serving cost, and treat every reported score as the authors' own until someone reproduces it.

If your problem is "both" — a face that converses — neither of these is the product, and the gap between them is where the next twelve months of this category will be fought. The renderer that streams, driven by a listener that decides, at a price you can model: nobody has shipped that yet. When someone does, this page will need rewriting.