Thinking

Ollama now supports enabling or disabling thinking, letting users choose a model's thinking behavior for different applications and use cases. When thinking is enabled, output separates the model's thinking from its output; when disabled, the model outputs content directly without thinking. DeepSeek R1 and Qwen 3 support thinking, and Ollama states more models will be added.

In the CLI, thinking is enabled by default. It can be disabled with /set nothink followed by the prompt in interactive sessions, or enabled with /set think. Command-line flags include --think to enable and --think=false to disable. A --hidethinking command is available for scripting, which Ollama says helps users who want to use thinking models but only see the answer.

Both the generate API (/api/generate) and chat API (/api/chat) have been updated to support thinking through a new think parameter that can be set to true or false. With think set to true, output separates the model's thinking from the model's output, which Ollama says can help craft application experiences such as animating the thinking process via a graphical interface or giving NPCs in games a thinking bubble before output. With think set to false, the model does not think and directs output.

The Python and JavaScript libraries have been updated, with installation via pip install ollama and npm i ollama. Examples in both libraries show passing think=True or think: true, and printing response.message.thinking and response.message.content. A JavaScript streaming example handles chunk.message.thinking and chunk.message.content. Ollama directs users to download the latest version and points to its GitHub for reference.