Examples

Pick a task below. These examples use the CUDA installation.

I want to…	Example
See the words as they are transcribed	Stream a transcript
Turn a sentence into a WAV file	Generate speech from text
Get a text answer about an audio clip	Ask a question about audio

Stream a transcript

What you need: The Qwen3-ASR-1.7B server and query_to_cars.wav from the Quickstart.

Start or reuse the server: Keep the Quickstart server running on port 8000.

Send a request: From the terminal where you downloaded the clip, upload it with stream=true. curl -N prints events as they arrive.

curl -fsS -N http://localhost:8000/v1/audio/transcriptions \
  -F model=Qwen/Qwen3-ASR-1.7B \
  -F file=@query_to_cars.wav \
  -F response_format=json \
  -F stream=true

Expected result: The server emits transcript.text.delta events as it decodes, then transcript.text.done with the full transcript, followed by data: [DONE]. The complete file is uploaded before decoding starts. For live audio and more options, see the Qwen3-ASR Cookbook.

Generate speech from text

What you need: A CUDA GPU, the Qwen3-TTS prerequisites, and a source checkout for the CustomVoice config. You do not need a reference recording.

Start or reuse the server: Stop the Quickstart server before reusing port 8000. From the root of the checkout, start Qwen3-TTS CustomVoice:

sgl-omni serve \
  --model-path Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
  --config examples/configs/qwen3_tts_1_7b_customvoice.yaml \
  --port 8000

Send a request: Ask the built-in speaker Ryan to read a sentence:

curl -fsS http://localhost:8000/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
    "input": "Hello from SGLang-Omni.",
    "voice": "Ryan",
    "language": "English",
    "response_format": "wav"
  }' \
  -o output.wav

Expected result: Play output.wav to hear the sentence. See TTS Model Usage for more speech requests.

Ask a question about audio

What you need: Qwen3-Omni-30B-A3B-Instruct and one CUDA GPU with enough memory for the BF16 thinker, such as an H200. The server fetches the Quickstart clip from GitHub. For other hardware and precision options, see the Qwen3-Omni Cookbook.

Start or reuse the server: Stop the previous server if you are reusing its GPU. Start a separate Qwen3-Omni server on port 8008:

sgl-omni serve \
  --model-path Qwen/Qwen3-Omni-30B-A3B-Instruct \
  --model-name qwen3-omni \
  --text-only \
  --allowed-media-domain raw.githubusercontent.com \
  --port 8008

If you are using the Quickstart Docker setup, run the server in that container.

Send a request: Ask the model to identify the spoken question:

curl -fsS http://localhost:8008/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "qwen3-omni",
    "messages": [{"role": "user", "content": "What question is spoken in this clip?"}],
    "audios": ["https://raw.githubusercontent.com/sgl-project/sglang-omni/7fc1a31850efbc1138e1895b5a3bcbeecf3a8481/tests/data/query_to_cars.wav"],
    "modalities": ["text"],
    "max_tokens": 128
  }'

Expected result: Check choices[0].message.content for a text answer that identifies the question about the number of cars in the picture. See Omni Model Usage for more request options.