ByteDance Seed has released SeedRealtime, an audio-visual full-duplex large language model built to participate in a live interaction rather than wait for a neatly packaged prompt. It processes continuous audio, video and text streams together, then decides whether to listen, respond or take an action as the situation changes.
That design targets a familiar weakness in real-time assistants. Many voice systems are still cascades: speech is transcribed, passed to a language model and converted back into speech, while any camera input follows another path. The hand-offs can introduce delay and strip away cues such as who is speaking, what they are looking at and whether an interruption changes the meaning of a request.
One conversation, not three separate pipelines
Seed describes SeedRealtime as a native, unified model for audio, video and text. Its first claimed breakthrough is joint audiovisual and temporal understanding. In practical terms, the model is intended to keep track of a changing scene and a changing conversation at the same time, rather than treating each frame or spoken turn as an isolated request.
The company demonstrates this with a noisy group discussion in which the system needs to associate voices with visible speakers. Other examples combine speech with what the camera can see: translating a restaurant menu in context, identifying an object in a museum, helping correct steps while someone uses a coffee machine, and following a requested section in a research paper.
These are demonstrations rather than independent evaluations, but they make the product direction clear. SeedRealtime is aimed at assistants that remain present during an activity and use the visual scene to resolve references such as “this”, “that person” or “the next step”.
Timing becomes part of the model’s job
Full-duplex interaction means the model can listen while it speaks and handle interruptions without forcing the user through rigid turn-taking. Seed says the system has been trained for natural conversational timing and interference resistance, including background speech and environmental noise.
The second capability is proactive interaction. Instead of replying only after an explicit command, SeedRealtime can identify a relevant moment and offer a reminder or invoke a tool. The examples include noticing an object during a museum visit and providing guidance when a user makes a mistake with a machine. That behaviour could make an assistant more useful, but it also raises product-design questions about consent, notification frequency and when silence is the better response.
Seed reports that human evaluation against cascaded systems cut conversational-pacing problems by half and improved the completion rate for smooth conversations. Those figures are vendor-reported, and the announcement does not provide enough methodological detail to compare them directly with other real-time multimodal systems.
The difficult part will be operating it responsibly
A continuously listening and watching assistant creates a different deployment problem from a text chatbot. Teams will need clear controls for camera and microphone state, retention, bystander data, tool permissions and the conditions under which the model may act proactively. The source does not spell out enterprise governance, regional availability, pricing or an API route.
Seed says the model has been fully rolled out, but the announcement is stronger on capability demonstrations than on commercial access details. Buyers and developers will therefore need to confirm how the model can be used, what data-processing terms apply and whether the released service supports the latency and concurrency required for their setting.
The roadmap includes lower latency, more natural timing, stronger decisions about when to intervene, better handling of multi-person environments and broader tool-driven action. Those priorities acknowledge that a convincing live assistant depends on more than model accuracy: it must know what is happening, respond at the right moment and avoid becoming an unwanted participant.