Unlock AI Reasoning: How Ollama’s New ‘Thinking’ Feature Works

Ollama v0.9.0 introduces a “thinking” control that allows users to control the AI model’s inference process, providing a new experience for AI application development.

What is the Thinking feature?

The thinking feature lets the AI model display its internal reasoning before giving the final answer, essentially “thinking out loud” so users can see step‑by‑step analysis.

Models that support Thinking

DeepSeek R1 – a powerful open‑source model

Qwen 3 – Alibaba’s multilingual large model

More models are being added gradually

CLI usage

Basic control commands

Enable thinking:

<code># Enable with parameter
ollama run deepseek-r1 --think "9.9和9.11哪个更大？"

# In interactive mode
/set think</code>

Disable thinking:

<code># Disable with parameter
ollama run deepseek-r1 --think= false "快速计算10+23"

# In interactive mode
/set nothink</code>

Scripting usage

To use a thinking model in a script but only see the final result, add the

--hidethinking

flag:

<code>ollama run deepseek-r1:8b --hidethinking "草莓这个词里有几个r？"</code>

API call example

REST API

Ollama’s API fully supports the thinking feature via the

think

parameter:

<code>curl http://localhost:11434/api/chat -d '
{
  "model": "deepseek-r1",
  "messages": [
    {
      "role": "user",
      "content": "解释一下量子纠缠的原理"
    }
  ],
  "think": true,
  "stream": false
}'
</code>

Response format includes a

thinking

field that contains the model’s step‑by‑step reasoning.

Java integration libraries

Current mainstream Java AI libraries do not yet support the thinking flag:

Spring AI – does not support the

think

property

LangChain4j – does not support the

think

property

For Java projects, it is recommended to call the Ollama REST API directly.

Modelfile advanced configuration

Seamless thinking chain control for Qwen 3

Qwen 3 can be instructed to skip deep reasoning by embedding the

/nothink

keyword in the prompt. Using the Modelfile

TEMPLATE

feature, this can be added automatically without changing backend code.

<code>FROM qwen3:latest
TEMPLATE ""
...
{{- if eq .Role "user" }}
<|im_start|>user
/nothink {{ .Content }}<|im_end|> # change this line
{{ else if eq .Role "assistant" }}
<|im_start|>assistant
...
""
</code>

Creating and using the model:

<code>ollama create qwen3-fast -f ./Modelfile
ollama run qwen3-fast "解释一下机器学习的基本概念"
</code>

The created model automatically prefixes each user input with

/nothink

, enabling a default fast‑response mode.

Reference resources

Ollama official blog: https://ollama.com/blog/thinking

Ollama documentation: https://ollama.com/docs