Support for Ollama's Thinking/Reasoning Models with think=True Parameter

[yuv-c]

Ollama recently introduced support for reasoning models with a think=True parameter that enables chain-of-thought reasoning (see Ollama's thinking blog post). When enabled, the model provides both its reasoning process (thinking field) and the final answer (content field).

I'm trying to use this feature with instructor in JSON mode, but I'm unable to access the thinking output.

Current Setup

import instructor
from pydantic import BaseModel

class QueryPlan(BaseModel):
    steps: list[str]
    explanation: str

client = instructor.from_provider(
    "ollama/qwen3:14b",
    mode=instructor.Mode.JSON,
)

# Trying to pass think=True via extra_body
response = client.chat.completions.create(
    model="qwen3:14b",
    messages=[{"role": "user", "content": "Create a plan for X"}],
    response_model=QueryPlan,
    temperature=0.2,
    extra_body={"think": True}  # <-- Attempting to enable thinking
)

Questions

1. Is the think parameter in extra_body actually passed to the Ollama API?

   • When I pass extra_body={"think": True}, does instructor forward this to Ollama's API?

2. Is thinking supported in JSON mode (instructor.Mode.JSON)?

   • Since the thinking field isn't part of my Pydantic schema, would it be stripped out even if Ollama returns it?
   • Is there a way to access the raw response that includes the thinking field?

3. Are there other modes that would preserve the thinking output?

   • Would instructor.Mode.TOOLS or another mode better support capturing both structured output and thinking?

4. How can I access both the thinking process and the validated response?

   • Ideally, I'd like to get both:
     • The validated Pydantic model (my QueryPlan)
     • The thinking/reasoning text for logging/debugging

Environment

• instructor version: 1.11.3
• ollama version: (latest)
• Python: 3.13

Related

• Ollama thinking documentation: https://ollama.com/blog/thinking

Thank you.