"/no_think" control is unstable
When passing "/no_think", "/no_thinking" or similar parameters into the system prompt, qwen3 generates no thoughts, as expected.
However!
The next conversation turn DOES include a thinking section.
so a conversation might look like this:
[
{"role": "system", "content": "/no_think"},
{"role": "user", "content": "hiii my sweetie qwen!"},
{"role": "assistant", "content": "<think>\n\n</think>\nhi! how may i assist u today?"},
{"role": "user", "content": "well, how do you do?"},
{"role": "assistant", "content": "<think>\nOkay, the user said \"blablabla\". Let's break this down step by step [...]\n</think>\nim doin fine, thanks for askin"}
]
in the example, the first reponse was correctly without thought content, but the latter did have thinking content.
Here an instance of this happening in ollama on qwen3:4b (Q4K)
I was dealing with stubborn thinking enabled despite specifically setting it to disabled in a roleplay scenario where I just wanted to see the direct response and nothing else. What I ended up doing was that I was prepending that <think>\n\n</think>\n into AI's response and it started giving me the direct answers. It's not ideal, but it works.