Qwen API: setup, endpoints and examples

2. Send a Python request

Install the compatible SDK in your Python environment:

python -m pip install --upgrade openai

After setting both environment variables, run this script. It selects qwen3.7-plus, disables thinking for a short text task and sets an output ceiling.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["DASHSCOPE_API_KEY"],
    base_url=os.environ["QWEN_BASE_URL"],
    timeout=120.0,
    max_retries=0,
)

response = client.chat.completions.create(
    model="qwen3.7-plus",
    messages=[{
        "role": "user",
        "content": "Explain a context window in two short sentences."
    }],
    max_completion_tokens=1024,
    extra_body={"enable_thinking": False},
)

choice = response.choices[0]
print(choice.message.content or "No final answer returned.")
print("Finish reason:", choice.finish_reason)
print("Usage:", response.usage)

enable_thinking is a provider-specific option. In the Python SDK it belongs in extra_body; in direct HTTP JSON it belongs at the top level. The current reference supports max_completion_tokens for this model family and defines it as the combined thinking-and-answer ceiling. Here thinking is disabled.

The example stops automatic retries so a first diagnostic call is easier to interpret. An output ceiling limits generation, not input cost or every possible application expense. Check the returned usage and your account records.

Parameter source: official Chat Completions reference. Check model-specific parameter support before switching to a different model.

3. Send the request with cURL

This uses Bash-style quoting. Both environment variables must already be set. QWEN_BASE_URL is the base ending in /v1, without the chat-completions path.

curl --fail-with-body --max-time 120 \
  "$QWEN_BASE_URL/chat/completions" \
  -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "qwen3.7-plus",
    "messages": [{
      "role": "user",
      "content": "Explain a context window in two short sentences."
    }],
    "enable_thinking": false,
    "max_completion_tokens": 1024
  }'

Read the answer from choices[0].message.content. Inspect finish_reason before treating the task as complete. A value of length indicates that generation reached its output limit; revise the prompt or budget before retrying.

Example status: documentation-based examples, reviewed for syntax and configuration. No authenticated live inference test or performance benchmark was performed by ChinaAI for this guide.

Before connecting this to a website

Keep the provider key on the server. Add visitor access controls, input limits, usage limits and spend monitoring before exposing a public chat form. An API key embedded in a browser page can be copied and used outside your site.

For temporary failures, use a request queue and bounded backoff. A timeout is not proof that no inference was performed: reconcile provider usage before repeating expensive calls. Log diagnostic identifiers without copying secrets or unnecessary user content into logs.

Add streaming only after a simple request succeeds. Your interface should handle an interrupted stream and show a failed request clearly. For structured data, use the documented output mode and validate the returned object before relying on it. For tools, validate arguments and retain control over any action your application executes.

See the official streaming guide and API interface index. Image generation, audio and local model serving need their own setup.