Real-time voice models have long faced a single-choice question: to push first-response latency into the millisecond range, you usually have to cut reasoning depth. Most products can only pick one side — do customer service, give up empathy; do companionship, don't expect tool-calling. In May this year, Alibaba's Qwen-Audio Preview topped the Artificial Analysis voice-reasoning leaderboard at 97.6%, but how to combine "fast" and "smart" in a large-scale real-time scenario had no official end-to-end solution. Qwen-Audio-3.0-Realtime, released on July 15, fills that gap. The core is On-Policy Distillation: while the voice model autoregressively generates output, a larger text LLM scores and corrects in real time, decoupling the "brain that thinks" and the "mouth that speaks" within the same forward pass during training. With four teachers — spoken-language preference, general reasoning, Agentic, audio understanding — the model simultaneously improves on IQ, empathy, Agent invocation, and duplex fluency, split into a more-capable Plus and a faster Flash. Even more interesting is the Agent dimension. Qwen-Audio-3.0-Realtime no longer needs explicit instructions to trigger tools, and call results are automatically deposited into conversation memory — meaning the voice side has, for the first time, the same FunctionCall experience as text LLMs, and natively supports the MCP protocol for connecting to external APIs and knowledge bases. For empathy and duplex, a "multimodal-perception duplex-control" sub-model is introduced, using audio signals, semantics, and voiceprint together to decide whether to interrupt or switch speakers. Around WAIC, every vendor is racing the voice Agent. Compared with TTS+ASR+LLM three-stage assembly, an end-to-end voice LLM is the truly reusable voice-Agent foundation. Alibaba's step here combines real-timeness, reasoning depth, and Agent capability into one distillation framework, and directly hangs on the MCP ecosystem — the cadence is not slow.