ByteDance Seed Releases Native Audio-Video Full-Duplex Model SeedRealtime

DeepFlow TechFlow reports that on August 10, ByteDance's Seed team launched SeedRealtime, a native audio-video full-duplex large model, now available on the Doubao App. This model integrates audio, video, and text into an end-to-end architecture, simultaneously processing perception, understanding, decision-making, and expression, while internalizing turn-switching within the model itself, replacing external voice activity detectors. Internal evaluations at ByteDance indicate that rhythm-related issues have been halved compared to cascade architectures. However, no technical report, parameter scale, open-source weights, or API has been released.