Quickstart

Start Qwen3-ASR-1.7B on one CUDA GPU and transcribe a short audio clip.

Install and start

Use Linux with a compatible NVIDIA GPU and driver. Choose Docker or uv below; see CUDA Installation for the requirements and other platforms for non-CUDA setups.

Docker (recommended): Follow Docker installation to pull the image, start the container, and install SGLang-Omni inside it. Run the launch command below in that container.

uv: With the CUDA prerequisites installed, create a Python 3.12 environment on the host:

uv venv .venv -p 3.12
source .venv/bin/activate
uv pip install --prerelease=allow "sglang-omni==0.2.0"

From either environment, start the server:

CUDA_VISIBLE_DEVICES=0 sgl-omni serve \
  --model-path Qwen/Qwen3-ASR-1.7B \
  --model-name Qwen/Qwen3-ASR-1.7B \
  --host 127.0.0.1 \
  --port 8000

The first launch downloads the checkpoint. Wait for the server to finish starting. If startup stalls, check the server log and run curl -fsS http://localhost:8000/health.

Send a request

In a second terminal on the same host, download the sample WAV used in the Qwen3-ASR Cookbook and upload it:

curl -fL -o query_to_cars.wav \
  https://raw.githubusercontent.com/sgl-project/sglang-omni/7fc1a31850efbc1138e1895b5a3bcbeecf3a8481/tests/data/query_to_cars.wav

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

Check the JSON response’s text field for the spoken question, “How many cars are there in the picture?”

Keep the server running to try streaming transcription.