A lightweight model that speaks while listening
"qwen-audio-3.1-realtime-flash," a lightweight version of a real-time voice conversation model, was listed on Oct. 27 (the night of the same day, Japan time) on "Model Studio," the developer service of China's Alibaba's cloud. @aitrackerbot, an X account that tracks new model releases, spotted the listing and reported it.
It supports "full-duplex" conversation, speaking while listening to the user's voice and responding even when interrupted midway. It accepts audio and text and responds in audio and text. It supports voice cloning, calling external tools and web search. Its context window is about 260,000 tokens, and it retains up to 50 turns of conversation history and a total of 300 seconds of audio.
| Type | Input | Output |
|---|---|---|
| Text | 0.23 | 0.70 |
| Audio | 0.93 | 1.87 |
One of September's audio model family
Alibaba on Sept. 23 released "Qwen-Audio-3.1," a family of five models covering speech recognition, text-to-speech and real-time conversation. The real-time model is said to respond slowly and with empathy when it senses that the user is feeling down. At that time, Alibaba cut prices sharply, by about 85% for the real-time model and by up to 95% for speech recognition.
For this lightweight version, Alibaba has not given performance figures or explained how it differs from the standard version. The supported languages are also not clear.
