By default, the template for qwen35/qwen35moe models discards thinking that occurs before the most recent user message. This is in accord with the default behaviour of the Jinja template for the model. However, the template also supports preserving earlier thinking, which is not supported in ollama due to the inability to pass in random parameters.

This template adds two roles, preserve_thinking and enable_thinking, mirroring the control variables in the Jinja template. These can be used to control thinking and preservation by sending messages rather than setting variables.

$ curl -s http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "frob/qwen3.6:35b-reflection",
    "messages": [
      { "role": "preserve_thinking", "content": "true" },
      { "role": "user", "content": "Choose a fruit but just reply with ok." },
      { "role": "assistant",
        "content": "ok", 
        "reasoning": "I selected avocado. Remember this for later if asked."
      },
      { "role": "user", "content": "Which fruit did you pick?" }
    ]
  }' | jq .choices[0].message
{
  "role": "assistant",
  "content": "Avocado"
}