What the model is
Alibaba Cloud Model Studio / DashScope qwen-audio-3.0-realtime-plus (QWEN_AUDIO_MODEL). Frontend-only speech-to-speech; the harness does not attach an ACP coding-agent backend. The scored deploy uses a Singapore (ap-southeast-1) workspace. The model is offered in Beijing and Singapore only.
How the harness was built
voice-agent-harnesses/qwen/ is a DashScope Realtime WebSocket client. The socket URL is built from QWEN_WORKSPACE_ID and QWEN_REGION, or from QWEN_WS_URL. Industry tools POST to the state API. Session tools hang up locally. Tools are declared in the nested Model Studio shape {type: function, function: {name, description, parameters}}.
Speak-first seeds a user conversation.item.create, then response.create. Qwen rejects a bare create on an empty conversation. Greeting text stays in the pack. Audio is Qwen-Audio PCM 16 kHz in / 24 kHz out, bridged to CHIRP 16 kHz pcm_s16le. Barge-in is Qwen server_vad.
Multi-agent assumptions
Handoffs are soft session.update on the same WebSocket, history stays. That matches the Grok pattern: the industry DAG swaps instructions and tools in place. Today's date is injected. infer_schedule_appointment recovers verbal booking confirms without a function-call event.
Runtime
server_vad defaults: threshold 0.5, silence_duration_ms 800 (QWEN_VAD_*). Session.update wait timeout: 60 s. Tool POST timeout: 30 s. Default voice longanqian, honored on the first session.update only. Local CHIRP port 8769 (8765 in-pod). Kubernetes resources are the default CHIRP box.
Deployment
CHIRP family on Kubernetes: run.py --harness qwen/audio-realtime --apply. Needs DASHSCOPE_API_KEY or QWEN_API_KEY, plus workspace id or QWEN_WS_URL, on mivas-secrets. The deployment assigns a dedicated WebSocket hostname to each harness and industry pair.
Source: mivas-bench, run.py, k8s/deployment.yaml.