Alibaba‘s Qwen team has launched Qwen-Audio-3.1-Realtime, a new full-duplex voice model designed to reason, call tools, and manage conversation timing independently. According to reporting by MarkTechPost, the model is available now as an API through QwenCloud.
Capabilities and Performance
The model introduces several quantitative improvements in conversational audio tasks. When evaluated on a tau-Voice adaptation, task success increased from 78.4 percent to 82.0 percent. Additionally, instances of the model incorrectly replying to background speech dropped significantly from 73 percent down to 13 percent.
Unknowns and Open Questions
Despite the release, the announcement did not specify the exact architecture size, training dataset composition, or hardware requirements needed to run the model locally. It remains unclear whether Alibaba plans to release open-weight versions of Qwen-Audio-3.1-Realtime for self-hosting, or if access will remain strictly limited to the QwenCloud API.
Developers will need to test the QwenCloud API directly to evaluate latency and integration capabilities for real-time voice applications.
Lead image: Towfiqu barbhuiya / Pexels
