Alibaba's Qwen team releases Qwen-Audio-3.1-Realtime full-duplex voice model on QwenCloud

A five-model audio stack led by a full-duplex realtime model

Alibaba's Qwen team released Qwen-Audio-3.1, a five-model audio stack covering automatic speech recognition, text-to-speech and realtime interaction. The centerpiece is Qwen-Audio-3.1-Realtime, a full-duplex speech model built for voice agents that call tools. It is offered as a managed API on QwenCloud over WebSocket, with no open weights announced. Qwen also cut prices across the stack: about 85% on Realtime, about 70% on TTS, and up to 95% on ASR.

Pricing, limits and deployment shape on QwenCloud

The QwenCloud model page lists text and audio as both input and output, with a 262K-token context (245K max input, 16K max output) and default limits of 60 requests and 100K tokens per minute. Audio input is priced at $6.4 per 1M tokens and text input at $0.8 per 1M tokens, while text and audio output are priced at $24 per 1M tokens, with text output not charged. The release supports function calling, web search, structured outputs, context caching and fine-tuning. A companion model, Qwen-Audio-3.1-ASR-Flash-Filetrans, targets offline long-audio transcription with hot words, speaker separation, punctuation and multilingual plus Chinese-dialect recognition, priced at $0.15 per 1M input tokens and $0.47 per 1M output tokens.

Two-model architecture trained to think, act and decide when to speak

The realtime system runs two models sharing one Audio Encoder and LLM design. A full-duplex decision model predicts whether to keep listening, speak, stop or resume, while a speech-to-text model drafts response content. A context-aware voice renderer then streams speech conditioned on conversation history, voice cues and acoustic context. Training is organized into three layers. The "Think" layer re-anchors the audio model to its source text LLM via million-hour-scale paired data and on-policy multimodal distillation, then merges empathy, pragmatic-intent and acoustic-scene experts with GRPO. The "Act" layer places each domain in an executable environment with a tool pool, a stateful JSON database and a natural-language business policy; rewards are issued at the dialogue, milestone and turn level, and a redundant-query penalty drove mean queries per search call down from 4.37 to 1.05. The "Speak and Coordinate" layer governs when and how to respond.

Measured gains and trade-offs

On the team's Full-Duplex-Bench, replies to people talking to someone else fell from 0.13 to 0.03 between versions, and the filler rate on v3.0 dropped from 0.7590 to 0.2960. Task success rose to 82.0%, up from 78.4%, while responses to background speech fell from 73% to 13%. Trigger F1 slipped from 60.87% to 58.61%, a trade-off Qwen flagged alongside residual unwanted responses after interruptions.

Distribution and access

Qwen-Audio-3.1-Realtime-Plus is live on QwenCloud over WebSocket as a managed API, with no accompanying open-weights release disclosed. Independent press coverage on the same day corroborated the launch, performance figures and the WebSocket-based deployment, while earlier reporting added detail on the on-policy distillation pipeline and search-training rewards.