SeedRealtime Technical Breakdown: How Full-Modal Full-Duplex Works

![Image 2: 向阳乔木](https://www.bestblogs.dev/en/tweets?sourceId=SOURCE_50f62a) 向阳乔木

@vista8

SeedRealtime，一个能看懂画面、理解手势、识别物品的全模态全双工交互模型。

现有主流音视频实时交互，要么是多模块级联（ASR 转文字、VLM 理解画面、TTS 合成语音），延迟层层叠加，信息逐级损耗。要么号称端到端，但轮次判断仍然依赖外置 VAD 模块，本质上还是一问一答的半双工逻辑。

用户说话的时候，模型在等。模型说话的时候，用户也在等。

SeedRealtime 将声音、画面、时序与表达纳入到同一个模型实时决策，听、看、理解、推理、说话全在一个模型里。在连续音视频流上，感知、理解、决策与表达同步进行，使听到的和看到的能够参与每次的实时判断。

比如当用户说"这个怎么弄"，"这个"是什么？在纯语音系统，问题无解。

在 SeedRealtime 里，模型会结合当前画面、用户的手势、视线方向和历史操作，判断"这个"的具体指向。当遇到同音词或语音模糊时，也能借助视觉场景消除歧义。

这里可以看我的实测，手指着天龙功放按键，模型理解回答，很自然。

据说 OpenAI GPT Live 也只是音频全双工，SeedRealtime 则是全模态（视频、音频、文本）全双工，强啊！Show More

!Image 3: Video thumbnail

02:45

Aug 6, 2026, 8:14 AM View on X

2 Replies

0 Retweets

2 Likes

2,450 Views ![Image 4: 向阳乔木](https://www.bestblogs.dev/en/tweets?sourceid=50f62a) 向阳乔木 @vista8

One Sentence Summary

The author explains how SeedRealtime integrates audio, visuals, timing, and expression into one model for real-time decision-making, unlike cascaded or semi-duplex systems.

Summary

The author breaks down SeedRealtime's architecture, contrasting it with traditional cascaded systems (ASR $\rightarrow$ VLM $\rightarrow$ TTS) which suffer from latency and info loss, or semi-duplex systems relying on external VAD. SeedRealtime processes everything in one model, allowing it to understand references like "how do I do this?" by combining current visuals, gestures, and gaze. The author claims it is "full-modal" (video/audio/text) whereas GPT Live is primarily "audio" full-duplex.

AI Screening

Influence Score 3

Published Yesterday

Language

Chinese

SeedRealtime 技术解读：全模态全双工如何工作

SeedRealtime Technical Breakdown: How Full-Modal Full-Duplex Works

SeedRealtime Technical Breakdown: How Full-Modal Full-Duplex Works

One Sentence Summary

Summary

Tags

🤖 問 AI