In my experience with Qwen3.6 35B-A3B, disabling thinking made the model generate inaccurate replies. Ask it for the recipe of egg salad and it gives you the recipe of an omelette.

Did I miss something, is it possible to have that model be reliable without thinking?