A real-time duplex speech model with text and audio input and output, function calling, web search, and voice cloning.
Inference service provider
Alibaba Cloud Model Studio.
Model capabilities
| Capability | Details |
|---|---|
| Input modalities | Text, audio |
| Output modalities | Text, audio |
| Function calling | Supported |
| Web search | Supported |
| Voice cloning | Supported |
| Protocols | WebSocket, WebRTC, AOQ |
| Default voice | longanqian_v3.1 |
| Thinking mode | enable_thinking is not supported |
| Interaction modes | server_vad, smart_turn, smart_turn_v2 (recommended), push-to-talk |
| Speaker enhancement | Supported in smart_turn and smart_turn_v2 modes |
enable_search. Web search and function calling can be enabled together.
Context limits
| Parameter | Limit (tokens) |
|---|---|
| Context window | 262,144 |
| Max input length | 245,760 |
| Max output length | 16,384 |
| Audio history in context | Maximum |
|---|---|
| Max turns | 50 |
| Cumulative audio duration | 300 seconds |
Pricing
Text-only responses are billed at the text output rate. For responses containing both text and audio, only audio output is billed; the accompanying text has no separate output charge.
Prices are in USD per million tokens.
- Singapore
- China (Beijing)
| Model ID | Text input | Audio input | Text output | Audio output |
|---|---|---|---|---|
qwen-audio-3.1-realtime-flash | $0.23 | $0.93 | $0.7 | $1.87 |
Rate limits
WebSocket RPM counts connection requests, not model responses. There is no separate limit on concurrent connections.
- Singapore
- China (Beijing)
| Model ID | RPM | TPM | Concurrent WebSocket connections |
|---|---|---|---|
qwen-audio-3.1-realtime-flash | 60 | 100,000 | No separate limit |