Muse Realtime Avatar vs SeedRealtime: Two Models You Can Only Watch

Neither of these has a rate card. Muse Realtime Avatar, announced by Meta on September 23, 2026 at Meta Connect, has no API, no endpoint, no price and no date — it is a capability of Meta's Muse agent, demonstrated in a research post titled "Bringing Your Muse to Life." SeedRealtime, released by ByteDance Seed on August 5, 2026, has no API, no weights, no technical report and no parameter count — it exists inside ByteDance's own Doubao app, and ByteDance's post says only that it has been "fully rolled out." Two of the largest consumer AI companies in the world shipped audio-visual real-time systems within seven weeks of each other and neither one is purchasable. That is the comparison, and it is a more interesting one than the benchmark table nobody can build.

What separates them is the direction the video flows. SeedRealtime watches and listens and talks about what it sees — audio, video and text into one end-to-end network, with turn-taking moved from an external voice-activity detector into the model itself. Muse Realtime Avatar generates: you hand it a reference image, a photograph or an illustration or an object, and it renders that thing speaking and gesturing in continuous video for the length of the conversation. SeedRealtime's video is input. Muse Realtime Avatar's video is output. Both are described as full-duplex, both are audio-visual, and they are doing opposite jobs with the label.

What each one is, and where it lives

Six lines, and the last two are the ones that decide anything.

• Video — Muse Realtime Avatar generates it, at 448×768 portrait resolution and 25 frames per second, watermarked with a transparent overlay so a viewer can tell it is not a camera feed; SeedRealtime perceives it, streaming frames into the same network that handles the audio.

• The architectural bet — Muse Realtime Avatar shares one token stream between voice and video, so an audio-driven Diffusion Transformer consumes Muse Realtime Voice's speech tokens directly and nothing is lip-synced afterwards; SeedRealtime folds turn-taking into the model, so who speaks and when stops being an external voice-activity detector's decision.

• Where it runs — Muse Realtime Avatar is a capability inside Meta's Muse agent; SeedRealtime is inside ByteDance's Doubao app.

• Developer access — neither has an API. Neither publishes weights. There is no serverless option, no hosted endpoint, and no third-party platform carrying either one.

• Price — unpublished for both, because there is no commercial offer on either side.

• Independent evaluation — none for either system. Every number either company has published about its own model is first-party, and no outside evaluator has run either one.

Two evidence bases, both first-party, and one of them is much thinner

Here the comparison stops being symmetrical. Meta published four figures for Muse Realtime Avatar, all company-reported, measured against baselines Meta chose, none reproduced outside the company: 870 ms from the end of your turn to the first byte of a synchronized voice-and-video reply; 448×768 portrait video at 25 frames per second; 60× fewer model evaluations than its own baseline; and 12 concurrent real-time sessions on one NVIDIA GB200, an 8× capacity gain over Meta's two-step BF16 baseline from four-bit quantization-aware training and latency-aware dynamic batching. Meta also published a preference comparison against two commercial avatar systems — 78/22 against one, 88/12 against the other — and disclosed that on mannerism the result against one of them is not statistically distinguishable from parity. That disclosure is worth more than the wins; it is the kind of result most launch posts leave out.

SeedRealtime's published evidence is one end-to-end human evaluation. ByteDance says that against cascaded systems — a vision model wired to an ASR model wired to an LLM wired to a text-to-speech voice — SeedRealtime halves audio-visual conversational pacing problems, with fewer cut-offs, fewer false triggers and slower, more natural replies. What ByteDance has not published is a sample size, a methodology, an annotator count, a latency figure in milliseconds, a benchmark name, or a score. One secondary report cites a 12% gain in conversational fluency over the team's prior models. That is the entire public evidence base.

So the honest ranking of verifiability runs: neither has an independent run; Meta's is a set of measurements with stated baselines and a disclosed negative; ByteDance's is a single preference claim whose design is not described. Neither is reproducible, but only one of them is specific enough to argue with.

The move each one made

Both systems are answers to the same problem, and it is worth seeing that the two answers are different.

For a system that watches, the hard part is knowing when to speak. A model that is continuously receiving camera frames and audio has to decide, without an external trigger, whether the person in front of it has finished, whether it has been addressed, and whether the object that just entered the frame is relevant. That is the problem SeedRealtime moved inside the model, and ByteDance made the move across three modalities rather than two — a harder version of what Google, OpenAI and NVIDIA each did for audio alone. The demo behaviours follow from it: binding a voice to a face in a group conversation, speaking up unprompted when something enters the frame, guiding someone through a museum or a repair.

For a system that renders, the hard part is staying coherent. Generating video in short causal chunks means each chunk is conditioned on the last, and the failure mode is drift — by minute three, the character has quietly become someone else. Meta's rolling window of recent video latents is the answer to that, and the shared token stream is the answer to a different failure, the dubbed look you get when audio is generated first and a lip-sync model is run over it afterwards.

Neither move is available in the other system. SeedRealtime has no reference image to stay faithful to, because it is not rendering anyone. Muse Realtime Avatar has no external world to track, because its entire job is producing a face that matches a voice.

Why an uncallable model is still a routing question

Both of these systems delegate. Muse Realtime Avatar sits on top of Muse Realtime Voice, which is the conversational layer of an agent whose reasoning runs on Muse Spark — the model does the rendering, not the thinking. SeedRealtime's design keeps the conversation going while background work happens, which means the work that exceeds what it can do inline goes somewhere else and comes back. In both architectures the thing the user interacts with is a front end, and the intelligence is a separate text model behind it.

That separate text model is ordinary inference, and it is the part of the stack you can actually buy. On OrcaRouter it is one endpoint covering 196 models from 15 providers, at provider list price passed through with zero markup, with automatic failover when the model you chose errors or times out. We do not host SeedRealtime — it is ByteDance's, inside Doubao, and there is nothing to route. We do not host Muse Realtime Avatar either, and neither does anyone else. What we carry is the layer both systems hand their hard work to, which is the layer that changes when you want a different model doing the thinking.

What would change the answer

For SeedRealtime, the missing piece is not a feature — it is the evidence. A technical report with a parameter count, a benchmark suite and a described human evaluation would move it from "a demonstration inside an app" to something a team could reason about. ByteDance's post also links generic "Try Dola" and "Try in Playground" destinations without claiming weights or a documented API, which is the pattern of a product feature rather than a model release.

For Muse Realtime Avatar, the missing piece is commercial: a rate card, an endpoint, a date, and the region and age terms Meta has not fully spelled out. Meta's post notes that its demos do not all reflect avatars available in the Muse app, which is the sentence that should govern any plan built around this.

Until one of those changes, the comparison has an unusually short form. Both models are audio-visual, both are full-duplex, both are genuinely impressive engineering, and neither one can be called from a terminal by anyone reading this. The difference is what you would do with them if you could: one gives you a character that talks, and one gives you a system that notices.