ByteDance Seed launches SeedRealtime full-duplex model

ByteDance Seed announced SeedRealtime, a native audio-visual full-duplex LLM. It handles audio, video and text in one unified architecture and works in real time over continuous multimodal streams. The company says the model is "fully rolled out" and calls this the industry's first large-scale deployment of audio-visual full-duplex technology. It did not say which products use it.

Seed frames the launch against two existing approaches. Cascaded systems chain separate ASR, VLM and TTS modules, which adds latency and loses information between stages. Many end-to-end models still use an external voice-activity detector (VAD) to decide when each side speaks, so they stay essentially half-duplex. SeedRealtime instead runs perception, understanding, decision-making and speech in parallel inside one model.

The company highlights three capabilities:

• Joint audio-visual understanding: using what is on screen to resolve homophones and references such as "how do I do this."
• Proactive interaction: speaking up unprompted when the scene changes, and calling tools as part of its responses.
• Conversational timing: knowing when to chime in or pause, while ignoring bystander chatter and background noise.

The only quantitative claim comes from an end-to-end human evaluation. Seed says SeedRealtime halved conversational pacing problems compared with cascaded systems, such as cutting users off, responding slowly after a pause, or reacting to background noise. It also says the model completed more conversations smoothly. Seed did not publish the evaluation's size, method or baselines.

Most of the remaining evidence comes from demos. In one, the model matches names to faces and voices at a four-person dinner. In another, it reads a Chinese-only menu for an English-speaking diner. The post does not mention public benchmarks, open weights or API availability.