Controlling Kimi K2 thinking output?

Forum
1

I’m experimenting with Kimi K2 but can’t find a way to distinguish between the thoughts and the final output. Isn’t the standard to wap thoughts in <think></think> so ti can be filtered out?

I’ve tried instructing the model doing so but with limited success.

Has anyone a solution to filter thoughts and only retrieve final output?

2

Could you share your system / user prompt, or a screenshot of the playground (Playground - GroqCloud) where this is happening? I’d love to take a stab on it; I think we can do it through system prompting.