Structured output

With response_format the reply is JSON: any valid JSON object, or one that follows a schema you send.

Lanes

Both modes work on deepseek-v4.1-flash:cached, deepseek-v4.1-flash:standard and qwen3.8-27b:standard. In testing, each lane returned JSON even when the prompt asked for a poem.

Follow a schema

cURL

curl https://api.getworg.com/v1/chat/completions \
  -H "Authorization: Bearer $WORG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash:cached",
    "messages": [{"role": "user", "content": "Rate this review from 1 to 10: The delivery was late but the support team fixed it the same day."}],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "review_score",
        "strict": true,
        "schema": {
          "type": "object",
          "properties": {
            "score": {"type": "integer"},
            "sentiment": {"type": "string", "enum": ["negative", "mixed", "positive"]}
          },
          "required": ["score", "sentiment"],
          "additionalProperties": false
        }
      }
    },
    "reasoning_effort": "none",
    "max_tokens": 200
  }'

message.content

{ "score": 7, "sentiment": "mixed" }

Set "strict": true, list every property in required, and set "additionalProperties": false.
The JSON is a string in message.content. Parse it in your code.

Any JSON object

{"type": "json_object"} makes the reply valid JSON, but the model picks the keys. Name the keys you want in the prompt, or use a schema.

Check in your code

If finish_reason is length, the output was cut at max_tokens and the JSON is incomplete. Raise max_tokens and retry.
Validate the parsed object against your schema before you store or act on it.
Reasoning text, when on, comes in reasoning_content, not in content, so it does not break the JSON. Set "reasoning_effort": "none" when you only need the data.

Next: Tool calling Reasoning