SeedRealtime is ByteDance's native audio-visual full-duplex model, released on August 5, 2026 — SeedRealtime watches, listens and talks at the same time over a live video call, and it is already shipping to real users inside the Doubao app. ByteDance's announcement offers no API and no weights, and SeedRealtime is absent from both of the model marketplaces we checked on the day it launched.
What SeedRealtime is: one end-to-end model fusing audio, video and text, instead of chaining speech recognition, a vision model and text-to-speech together.
Released: August 5, 2026, by ByteDance's Seed team.
Where to try SeedRealtime: the Doubao app — update to the latest version and tap "call" (打电话) in the chat box.
The one hard number ByteDance published: human evaluation puts SeedRealtime's conversational pacing problems at half the rate of cascaded models.
Can you build on SeedRealtime? Not today. The announcement gives one access path — the Doubao app — and names no API or weights.
Last updated: August 2026.
What SeedRealtime's full-duplex design actually means
Full-duplex is a telephony word and it is the right one. Half-duplex is a walkie-talkie — one side transmits, the other waits — and that is the shape of a normal voice assistant. Full-duplex is a phone call: both ends stay open, and each side is continuously deciding whether to speak. SeedRealtime extends that from audio to audio and video, which ByteDance frames as "watch, listen, and speak" happening at once, over continuous streams rather than turns.
Why the usual way of building this falls short
Most real-time assistants, unlike SeedRealtime, are cascaded: ASR transcribes speech, a vision-language model looks at frames, TTS reads the answer out. ByteDance is blunt about the cost — chaining modules introduces "latency and information loss between stages." Every hand-off flattens nuance into text and throws the rest away.
The second failure mode is subtler, and it is the one that makes assistants feel robotic. Even genuinely end-to-end models usually still hand turn-taking to an external VAD (voice activity detection) — a component whose whole job is deciding "the human stopped, go now." A model with a VAD on the front is, in ByteDance's words, "essentially a half-duplex, one-question-one-answer interaction," however good the model behind it is.
SeedRealtime drops the VAD. Turn-taking becomes something SeedRealtime decides itself, continuously, from what it hears and what it sees.
The one number ByteDance published about SeedRealtime
There is exactly one quantitative claim in the SeedRealtime announcement, and it is worth reading precisely:
End-to-end human evaluation shows that, compared with cascaded models, SeedRealtime reduces audio-visual conversational pacing issues by half.
Note what SeedRealtime is being measured on: not accuracy, not benchmark scores — pacing. The failures named are:
- being cut off mid-sentence
- responding sluggishly after a pause
- being falsely triggered by background noise and bystander chatter
ByteDance also reports that the odds of getting through a conversation smoothly improved "significantly," without attaching a figure. No latency milliseconds were published for SeedRealtime, and no third-party benchmark exists yet — treat the rest as a vendor demo until someone outside ByteDance measures it.
What SeedRealtime can do, in three real scenes
ByteDance's post walks through seven demo scenarios:
- Dinner with four friends — matching names to faces, keeping each voice tied to a person
- A Sichuan restaurant — reading a Chinese-only menu for a foreign diner
- The Hebei Museum — speaking up when a named artefact enters frame
- An espresso machine — correcting technique unprompted
- The ResNet paper — stopping at a named section as pages flip
- Beijing Daxing Airport — ignoring bystander chatter
- A home English lesson — following a child's pointing over a parent's phone call
Three of them show what SeedRealtime's architecture buys.
Joint audio-visual understanding
A foreign diner faces a Chinese-only menu at a Sichuan restaurant. SeedRealtime reads the dishes off the scene, recommends them in English, and explains why there is no fish in yuxiang rousi and how century eggs are made. When a server sets down a plate and remarks that it "goes great with rice," SeedRealtime reads that line against the dish on screen and translates it.
The ordering is the point: the visual is never turned into text first and reasoned about after. Sound and image align inside one model, so an ambiguous phrase gets resolved by what is physically in frame — the same mechanism that lets SeedRealtime disambiguate homophones and work out what "this" refers to when a user points.
Proactive interaction
At the Hebei Museum, a visitor says: remind me when you see the gold-and-silver-inlaid bronze tiger-devouring-a-deer screen stand. The camera keeps moving, and when that piece pans into view, SeedRealtime speaks up unprompted — holding a task in context and monitoring a stream against it, rather than waiting to be asked. A second demo has SeedRealtime interrupting someone at an espresso machine when whole beans go into the portafilter ungrounded, then reading the crema colour and suggesting a 2–3 second shorter shot.
Knowing when not to speak
Beijing Daxing Airport, crowded and loud. A companion mentions "Old Li's flight" in passing chatter, and SeedRealtime does not take it as a cue to answer. When the user genuinely asks, the departure board has already scrolled off screen — so SeedRealtime answers from what it saw earlier, then goes online for the baggage carousel.
Restraint is the hard half. A model that can interject at any moment, without the judgement to know when not to, is worse company than one that waits its turn — which is why SeedRealtime's headline result is about pacing, not accuracy.
Can you use SeedRealtime?
For most people reading this in English, SeedRealtime is not usable directly, and not in your own product.
| SeedRealtime access | Status, checked 5 August 2026 |
|---|---|
| Announced | August 5, 2026, by ByteDance Seed |
| Access path ByteDance gives | The Doubao app, via the "call" feature |
| API mentioned in the announcement | None |
| Weights released | None announced |
| Listed on KIE or PoYo model catalogues | Not present in either (both catalogues re-fetched 5 Aug 2026: 102 and 98 models) |
| Published benchmarks | One: pacing issues halved vs cascaded, ByteDance's own human eval |
One correction, because it is spreading through English coverage: several write-ups say SeedRealtime shipped in Douyin. It did not. ByteDance's Chinese announcement reads 「目前，SeedRealtime 已在豆包 App 全量上线」 — fully rolled out in Doubao, its assistant app, not the short-video app. The access path: update Doubao, open the chat box, tap "call."
ByteDance describes SeedRealtime purely as an interaction model — nothing in the announcement claims image or video generation. That is a different branch of the Seed family, and parts of that branch you can build on today.
Where SeedRealtime sits in the Seed family
Here is the lineup with the column that decides whether you can build on any of it — whether the model shows up on the two aggregator catalogues most small products buy capacity through:
| Model | What it does | Announced | On KIE or PoYo (checked 5 Aug 2026) |
|---|---|---|---|
| SeedRealtime | Real-time audio-visual conversation | Aug 5, 2026 | No |
| Seeduplex | Real-time speech conversation | Apr 9, 2026 | No |
| Seed Audio 1.0 | Audio and sound generation | Jul 19, 2026 | No |
| Seed2.1 | General agent and coding model | Jun 23, 2026 | No |
| Seedance 2.5 | Video generation | Jul 30, 2026 | Yes |
| Seedream (4.0 – 5.0 Pro) | Image generation | Feb – Jul 2026 | Yes |
The split is clean: every conversational model in the family stays inside ByteDance's own apps, and every model that reaches outside is a generation model. Seedance for video and Seedream for images are the two that run in the generator on this site.
Seeduplex came first, four months earlier
On April 9, 2026, the same team released Seeduplex, a native full-duplex speech model built on a "listen while speaking" framework and pitched against ByteDance's older half-duplex Doubao voice model. It carried its own number — premature responses down 40% — and landed exactly the way SeedRealtime just did: "fully rolled out on the Doubao App," with the same claim to a first large-scale deployment.
April solved turn-taking for audio; SeedRealtime added the camera in August. Read in that order, SeedRealtime is the second half of a plan, not a one-off launch. The roadmap ByteDance published for SeedRealtime points the same way again:
- Lower end-to-end latency, and timing that reproduces real conversational habits — backchannelling, jumping in, pausing — rather than merely being faster
- More proactive perception, deciding on its own when to remind and when to add information
- Multi-person robustness — working out who is speaking, what they are looking at, and who to answer in a noisy room
- From "able to converse" to "able to act", by wiring tool calls into the conversation
ByteDance attached no dates to any of it.
SeedRealtime FAQ
What is SeedRealtime?
SeedRealtime is ByteDance's native audio-visual full-duplex large model, announced August 5, 2026. SeedRealtime handles audio, video and text in one end-to-end model, so it can watch, listen and speak at once during a live call instead of taking turns.
Is there a SeedRealtime API?
None has been announced. ByteDance's launch post names no SeedRealtime API and no weights release, and gives one access path: the Doubao app. Checked on launch day, SeedRealtime was also absent from the KIE and PoYo catalogues.
Which app is SeedRealtime in — Doubao or Douyin?
Doubao. Some English coverage says Douyin, but ByteDance's own Chinese announcement puts SeedRealtime in the Doubao app. Update Doubao, open a chat, tap "call."
If you only remember one thing: SeedRealtime is a conversation model that ByteDance runs itself, in its own app. On launch day there was no announced way to integrate it, and it is not a generator.
How is SeedRealtime different from a normal voice assistant?
A normal voice assistant is half-duplex: it waits for you to stop, usually via a separate voice-activity detector, then replies. SeedRealtime decides continuously whether to speak, from what it hears and what it sees — so it can interject, stay quiet through background chatter, or volunteer something when the scene changes.
Is SeedRealtime better than a cascaded ASR plus vision plus TTS setup?
On the one metric ByteDance published, yes: human evaluation put SeedRealtime's pacing problems at half the rate of cascaded models. That is ByteDance's own evaluation, with no independent benchmark yet — a vendor claim until someone reproduces it.
Can SeedRealtime generate images or video?
ByteDance presents SeedRealtime as a conversational model and claims no generation capability for it. Image and video generation come from different models in the same family — Seedream for images, Seedance for video — and those are the ones that turn up outside ByteDance's own apps.
What is Seeduplex, and how does it relate to SeedRealtime?
Seeduplex is the direct predecessor, released April 9, 2026: a native full-duplex model for speech only, which ByteDance also described as fully rolled out on the Doubao app. SeedRealtime is the version that adds vision.
Try the Seed models you can actually run
SeedRealtime is worth watching and you cannot build on it. The generation half of the family you can: Seedance 2.0 for video, Seedream 5.0 Pro for images, one shared balance, and what each costs shown before you generate.
Resources
- SeedRealtime Audio-Visual Full-Duplex LLM Released — ByteDance Seed, August 5, 2026 (primary source)
- SeedRealtime project page — ByteDance Seed
- Introducing Seed Full-Duplex Speech LLM — ByteDance Seed, April 9, 2026 (Seeduplex)
