Published: 6 August 2026
ByteDance’s Seed organisation launched SeedRealtime on 5 August 2026, a native audio-visual full-duplex large language model that processes continuous audio, video and text streams while listening and responding in real time. SeedRealtime fuses the three modalities in a single architecture rather than chaining separate speech recognition, vision-language and text-to-speech modules, and ByteDance reports that it halves audio-visual conversational pacing problems compared with cascaded systems. The model has been rolled out in the Doubao app.
What is SeedRealtime and how does it work?
SeedRealtime is built to watch, listen and speak over continuous streams at the same time. It combines audio, video and text in one unified architecture, which lets perception, understanding, decision-making and expression run in parallel. Conventional voice assistants hand off between a speech recognition module, a vision-language model and a text-to-speech module, and each handoff adds latency and loses context. SeedRealtime removes those handoffs.
The central shift is that the model decides conversational timing itself instead of relying on external voice-activity detection rules. It tracks scenes, speakers, pauses and background chatter to judge what matters and when to answer. Visual context does real work here: it can resolve homophones, connect words such as ‘this’ to a gesture or an earlier action, and retain information that has since moved off-screen.
The model can also speak without a new prompt. ByteDance gives two examples: it can raise a reminder when a requested object comes into view, and it can flag a mistake as it happens.
What can SeedRealtime do in practice?
ByteDance demonstrated the model in noisy, open settings rather than controlled ones. It matched names, faces and voices during a group dinner, interpreted dishes and speech in a restaurant, and issued a reminder when a requested museum object entered view.
In other examples it corrected an espresso-making mistake as it happened, spotted a requested section while the pages of a paper were being turned, ignored unrelated airport chatter, and followed a child’s pointing during an English lesson. In the airport example it went online to retrieve baggage-carousel information, which shows the model combining live perception with an external lookup.
Taken together, these are the situations that break cascaded assistants: several people talking, relevant visual detail that is never spoken aloud, and background speech that should not trigger a response.
SeedRealtime performance and what ByteDance has not disclosed
In end-to-end human evaluation, ByteDance says SeedRealtime cut audio-visual conversational pacing problems by half against cascaded models. Evaluators reported fewer cutoffs, fewer slow replies after pauses and fewer false triggers from nearby speech, while more conversations were completed smoothly.
The disclosure stops there. The announcement gives no sample sizes, no benchmark table and no exact figure for the completion gain, so the halving claim cannot currently be checked against a published methodology. There is also no parameter count and no architecture detail beyond the description of a unified audio, video and text design.
Seed states the model is fully rolled out but does not specify an API, supported regions, pricing or access requirements. Availability is through the Doubao app, where it is free and supports voice and video conversation.
Where SeedRealtime fits and what comes next
The release lands in the same week as Meta’s Muse Spark 1.2 and NVIDIA’s Alpamayo 2 Super, and it points in a different direction from both. Where those target coding and driving, SeedRealtime is aimed at continuous real-world interaction, and ByteDance frames it as a step toward omni-modal systems that observe, converse and act in changing environments.
Seed’s stated roadmap covers lower latency, finer timing for interruptions and backchannels, stronger speaker tracking in multi-person scenes, more proactive decisions, and tool-connected tasks such as lookups and bookings. The airport baggage example is an early instance of that last category.
For teams evaluating real-time multimodal assistants, the practical question is access. Without a documented API, pricing or regional availability, SeedRealtime is currently a consumer product in the Doubao app rather than something that can be built on.
SeedRealtime is live in the Doubao app, and the technical announcement on SeedRealtime is published on ByteDance Seed’s blog.
