Qwen-Audio-3.1-Realtime-Plus

Qwen-Audio-3.1-Realtime-Plus is an end-to-end, full-duplex speech-to-speech model from Alibaba, served through Alibaba Cloud Model Studio. Model Studio's model lifecycle page lists it as released on 20 September 2026 in both the Singapore (international) and China (Beijing) regions. Alibaba announced it publicly on 23 September 2026, and The Decoder reported the launch the same day, as part of Qwen Audio 3.1 with realtime prices cut by about 85 per cent. The context window and per-token prices below are Alibaba's own figures. It is intended for voice assistants, customer-service lines and AI companions.

The model takes streaming audio in and returns streaming speech and text over a persistent WebSocket connection; Alibaba also supports its AOQ protocol and WebRTC. Alibaba states a 262,144-token context window, function calling, web search (which cannot be enabled at the same time as function calling) and voice cloning. It keeps the integration protocol of Qwen-Audio-3.0-Realtime-Plus, so existing integrations change only the model name, and adds eight system voices to the existing set. Turn detection can be acoustic, semantic ('smart turn') or push-to-talk. Conversations are supported in eleven languages, including English, German, French, Spanish, Japanese and Korean, and in 21 Chinese varieties including Mandarin and Cantonese. History is kept for up to 50 turns (20 by default) or 300 seconds of cumulative audio.

Pricing is per million tokens. In Singapore, text input costs $0.80, audio input $6.40, text output $6.40 and audio output $24; mainland China rates are lower. Audio is counted at 12.5 tokens per second, and earlier turns are billed again as input on each new turn, so long sessions become progressively more expensive. A one-million-token free quota applies in Singapore only, for 90 days.

The model is proprietary: there are no downloadable weights and no published parameter count. Input audio must be 16 kHz mono PCM, and output is 24 kHz PCM.