Local models thinking mode visibility issue

[trozek24]

I'm using qwen3:14b, and I am new to using adk.
i can't find anyway to solve this issue of the model giving output of showing its chain-of-thoughtss

[surajksharma07]

@trozek24 — Qwen3 has extended thinking mode on by default, which is what's causing that reasoning output to show up.

Worth trying: add /no_think at the end of your agent's instruction string — it's a Qwen3 native flag that disables the chain-of-thought and should clean up the response.

Could you also share your ADK version (pip show google-adk) and Ollama version?

[trozek24]

ollama version is 0.21.0
adk, version 1.33.0

i tried /no_think , still the thinking is somehow leaking.
I also tried to use regex to remove the thinking tokens using regex which was adviced by gemini but to no avail

[surajksharma07]

@trozek24 With Ollama 0.21.0, Qwen3's thinking comes through as a separate reasoning_content field rather than embedded in the text — that's why regex won't catch it and ADK maps it to part.thought == True event parts instead.

The /no_think string in the instruction isn't handled as a control token at the Ollama API level.
The proper way to suppress it is to pass the think flag directly to Ollama via LiteLLM: LiteLlm(model="ollama_chat/qwen3:14b", extra_body={"think": False}).
If you'd rather keep thinking enabled but just hide it from your output, you can filter on the consumer side by skipping any event parts where part.thought == True before displaying the response.

Could you also confirm which interface you're using to display output — adk web, adk run, or a custom runner?
The Dev UI intentionally renders thought parts as a collapsible section, which might also explain what you're seeing!

[trozek24]

I'm currently using it on adk run for experimental purposes as I'm new to agents stuff. I'll later move on to ask web if i start building projects.

Apologise for my stupidity but were did you find this information, i tried to ask AI for answers but it couldn't and im too novice to search through those documents (and most of the time i dont find those docs), and tips you could suggest for explorinng and a begineer techy like me?