Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/QwenLM/Qwen3-ASR/llms.txt

Use this file to discover all available pages before exploring further.

Qwen3-ASR integrates natively with vLLM, giving you a production-ready, OpenAI-compatible HTTP server for speech recognition. Once the server is running you can call it with the OpenAI Python SDK, plain requests, or cURL — no custom client code required.

​
Installation

vLLM provides day-0 support for Qwen3-ASR. Use uv to install the nightly wheel along with the extra audio dependencies: 
uv venv
source .venv/bin/activate
uv pip install -U vllm --pre \
    --extra-index-url https://wheels.vllm.ai/nightly/cu129 \
    --extra-index-url https://download.pytorch.org/whl/cu129 \
    --index-strategy unsafe-best-match
uv pip install "vllm[audio]"   # additional audio dependencies
The qwen-asr-serve CLI command (described below) requires the vllm extra from the qwen-asr package. Install it with:
pip install -U "qwen-asr[vllm]"
Without this extra, qwen-asr-serve will raise an ImportError on startup.

​
Starting the Server

You have two ways to launch the inference server: Option 1 — qwen-asr-serve (recommended) qwen-asr-serve is a thin wrapper around vllm serve that automatically registers the Qwen3-ASR model architecture. It accepts every flag that vllm serve supports: 
qwen-asr-serve Qwen/Qwen3-ASR-1.7B \
  --gpu-memory-utilization 0.8 \
  --host 0.0.0.0 \
  --port 8000
Option 2 — direct vllm serve If you have already installed vLLM and registered the model separately, you can invoke vLLM directly: 
vllm serve Qwen/Qwen3-ASR-1.7B
Both commands expose the same OpenAI-compatible endpoints on the configured host and port.

​
Sending Requests

After the server is running, you can query it using the OpenAI SDK, the transcription API, or cURL. 
import base64
import httpx
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="Qwen/Qwen3-ASR-1.7B",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "audio_url",
                    "audio_url": {
                        "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav"
                    },
                }
            ],
        }
    ],
)

print(response.choices[0].message.content)

​
Parsing the Response

The raw model output encodes both the detected language and the transcription text in a structured format. Use parse_asr_output from the qwen_asr package to split them apart: 
from qwen_asr import parse_asr_output

# `content` is the string from response.choices[0].message.content
language, text = parse_asr_output(content)
print(language)   # e.g. "English"
print(text)       # the transcribed text

​
Offline Inference with vllm.LLM

For batch processing without an HTTP server, use vLLM’s LLM class directly. Wrap your script in a __main__ guard to avoid multiprocessing issues: 
from vllm import LLM, SamplingParams
from vllm.assets.audio import AudioAsset

# Initialize the LLM
llm = LLM(model="Qwen/Qwen3-ASR-1.7B")

# Load a bundled vLLM audio asset (or supply your own URL)
audio_asset = AudioAsset("winning_call")

conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "audio_url",
                "audio_url": {"url": audio_asset.url},
            }
        ],
    }
]

sampling_params = SamplingParams(temperature=0.01, max_tokens=256)

outputs = llm.chat(conversation, sampling_params=sampling_params)
print(outputs[0].outputs[0].text)
Always wrap vLLM offline inference code in if __name__ == "__main__": to prevent the spawn error that arises from Python’s multiprocessing model. See the vLLM Troubleshooting guide for details.

Build docs developers (and LLMs) love

Get started for freeTalk to us