Qwen3.5 2B burns all the output tokens while thinking

I am experimenting with the model and then model spends all its output tokens while thinking making no room left for final output. I have even set thinking budget, but still does not work, anybody has any workarounds or something I am missing?