Qwen

/

Qwen3.5-35B-A3B

like 1.4k

Follow

Qwen 80.7k

I’m using an OpenAI-compatible Chat Completions API endpoint with structured output (response_format = json_schema).

Expected behavior:

• The final JSON output should be returned in choices[0].message.content.
• Reasoning/thinking text should be separated (e.g., in reasoning_content) and never mixed into content.

Actual behavior:

• In some responses, reasoning text is injected directly into choices[0].message.content before or around the JSON payload.
• This breaks JSON parsing, even though response_format is set and schema is valid.

Context:

• API style: OpenAI Chat Completions
• Structured output: enabled via response_format: { type: "json_schema", ... }
• Thinking mode: enabled on the model side
• Parsing flow expects message.content to be strict JSON

Could you please clarify whether this is expected for thinking-enabled models, and if there is a reliable way to enforce strict JSON-only message.content while keeping reasoning in a separate field?
Thanks!