Structured output | Qwen Cloud

Structured output guarantees the model returns valid JSON. Set response_format to {"type": "json_object"} and include the word "JSON" in your prompt.

OpenAI compatible
DashScope

from openai import OpenAI
import os

client = OpenAI(
  api_key=os.getenv("DASHSCOPE_API_KEY"),
  base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

completion = client.chat.completions.create(
  model="qwen3.6-plus",
  messages=[
    {"role": "system", "content": "Extract the name and age. Return JSON."},
    {"role": "user", "content": "My name is Alex Brown, I am 34 years old."},
  ],
  response_format={"type": "json_object"},  # <-- force valid JSON output
)
print(completion.choices[0].message.content)

import os
import dashscope

dashscope.base_http_api_url = 'https://dashscope-intl.aliyuncs.com/api/v1'

response = dashscope.Generation.call(
  api_key=os.getenv('DASHSCOPE_API_KEY'),
  model="qwen3.6-plus",
  messages=[
    {"role": "system", "content": "Extract the name and age. Return JSON."},
    {"role": "user", "content": "My name is Alex Brown, I am 34 years old."},
  ],
  result_format='message',
  response_format={'type': 'json_object'},  # <-- force valid JSON output
)
print(response.output.choices[0].message.content)

Output:

{"name": "Alex Brown", "age": 34}

For more reliable results, describe the expected fields, types, required/optional status, and provide examples in your prompt.

Thinking mode workaround

Structured output is not supported in thinking mode. As a workaround, parse the thinking-mode output and fall back to a fast model for JSON repair.

Python

import json
from openai import OpenAI
import os

client = OpenAI(
  api_key=os.getenv("DASHSCOPE_API_KEY"),
  base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

# Step 1: Get output from thinking mode (streaming required)
completion = client.chat.completions.create(
  model="qwen3.6-plus",
  messages=[
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_input},
  ],
  extra_body={"enable_thinking": True},            # <-- no response_format here
  stream=True,
)
json_string = ""
for chunk in completion:
  if chunk.choices[0].delta.content:
    json_string += chunk.choices[0].delta.content

# Step 2: Parse, or repair with a fast model
try:
  result = json.loads(json_string)
except json.JSONDecodeError:
  repair = client.chat.completions.create(
    model="qwen3.5-flash",                       # <-- fast model fixes the JSON
    messages=[
      {"role": "system", "content": "Fix this to valid JSON."},
      {"role": "user", "content": json_string},
    ],
    response_format={"type": "json_object"},
  )
  result = json.loads(repair.choices[0].message.content)

Notes

Supported models: Qwen3.6, Qwen3.5, Qwen3, Qwen3-Coder, Qwen2.5, and legacy (Plus/Max/Flash/Turbo) — non-thinking mode. Also works with vision models (qwen3-vl-plus, etc.) when passing images or video as described in Image and video understanding.
Don't set max_tokens: Truncated output produces invalid JSON that breaks downstream parsing. Use the default (model's maximum output limit).
Validate output: JSON Object mode guarantees valid JSON but not schema conformance. Validate with jsonschema (Python), Ajv (JavaScript), or Everit (Java) before passing to downstream systems.

​Thinking mode workaround

​Notes

Thinking mode workaround

Notes