DuplexOmni

Model Description

This repository provides the DuplexOmni model weights. Both the Thinker and Talker components are included in the released checkpoint.

For source code and deployment details, see the DuplexOmni repository.

Known Issues

This model is still under active optimization and may exhibit the following issues:

• Unexpected silence during generation.
• Suboptimal speech quality or speech performance in certain cases.

Hardware Recommendation

For low-latency deployment, we recommend using at least 8 NVIDIA H20 GPUs. This recommendation reflects the current optimization level and may change as inference performance improves.

Warm-up Requirement

The model should be warmed up after deployment. Without warm-up, the first response packet can have significantly higher latency.

---

DuplexOmni

模型说明

本仓库提供 DuplexOmni 模型权重，其中同时包含 Thinker 和 Talker 两个组件。

源码及部署相关信息请参考 DuplexOmni 仓库。

已知问题

该模型目前仍在持续优化中，可能存在以下问题：

• 生成过程中出现意料之外的沉默。
• 部分情况下语音质量或语音表现不佳。

硬件建议

为获得较低的推理延迟，建议至少使用 8 张 NVIDIA H20 GPU 进行部署。该建议基于当前的优化水平，后续可能会随着推理性能优化而调整。

预热要求

模型部署并挂载后需要进行预热；否则，首个响应包的延迟可能会明显偏高。