Qwen Console

Live decode—tok/s · streaming

TTFT—time to first token

Output—tokens this reply

Backend load—running / queued

✦

Ask anything

Pick a model and choose the exact Qwen thinking effort.

Audio studio

Turn conversation, full-duplex talk, expressive speech, and songs use different purpose-built models.

Speak or upload release to send · slide up to cancel

checking

Listening model

Fast end-to-end English turn-taking. It waits for the complete clip; this is not full duplex.

Reply voice Tone Language

Persona / request optional

Microphone access needs HTTPS or localhost. A recording is capped at 30 seconds.

Drop an audio clipor choose WAV, MP3, OGG, WebM/Opus, M4A, MP4 · up to 32 MB

Response waiting for audio

Ready

Hold the button to speak, or upload a complete clip.

Throughput—tok/s · all models

Active—running requests

Pending—queued requests

Total calls—since start

Prefix cache—hit rate

Mean call—inference time

GPUs utilization & VRAM

Throughput tok/s · last 2 min

History from the audit log — survives reloads and restarts

MTP acceptance draft accept · last 2 min

Backends

Model	GPU	Run	Wait	tok/s	KV	Cache	MTP

Model endpoints what's serving right now

Model	Status	Engine	GPU	Endpoint (OpenAI /v1)	Ctx	Capabilities

Access points how to reach these models

Call History —

Time	Model	Src	Client	Status	Recv	Prompt	Out	TTFT	tok/s