Overview
SeedRealtime is a native audio-visual, full-duplex large language model for continuous real-time interaction. Its unified end-to-end architecture combines audio, video and text so that it can perceive, understand and respond while a conversation is ongoing. (seed.bytedance.com)
Capabilities and uses
- Jointly interprets visual, audio and timing information, including ambiguous speech and references to things in the scene.
- Tracks conversational timing, distinguishes speakers and filters some background interference.
- Can proactively comment on changes in a scene and incorporate tool calls into an interaction.
- Demonstrated uses include cross-language communication, explaining visual material, reminders during museum visits and guidance while operating equipment. (seed.bytedance.com)
Availability and limitations
ByteDance states that SeedRealtime has been rolled out in the Doubao app; its Chinese announcement directs users to the video-call interface. The announcement does not specify a standalone API model ID, context window, numeric latency, supported-language list or model weights. ByteDance identifies latency, conversational timing and robustness in complex multi-person environments as areas for further improvement. (seed.bytedance.com)
Pricing
As of November 16, 2026, a SeedRealtime-specific official input or output price could not be verified in either the mainland China Volcengine pricing documentation or the international BytePlus ModelArk pricing documentation. The billing unit is also unverified; prices for other audio or realtime models should not be substituted. (console.volcengine.com)
Release and official sources
- Release date: August 5, 2026. (seed.bytedance.com)
- Official English announcement
- Official Chinese announcement and Doubao access instructions
- Supplied official model page
- Pricing sources checked on November 16, 2026: Volcengine Model Pricing and BytePlus ModelArk Pricing.