For our bounding box detection tasks on documents, we found that gemini-flash-2.5 performs quite well, but only if thinking_budget=0.
If thinking_budget>0 the bounding boxes are much worse and sometimes quite far away from the actual object one is trying to detect.
Has anybody made similar observations?
1 Like
Hi @TheVoogler , Welcome to the forum.
Thanks for pointing out the issue. I have noticed the same thing, when thinking_budget is greater than 0, the model’s performance seems to drop.
I will be escalating this to the team.
1 Like
It would really be a great tool if the bouding boxes were correct when doing anything other than basic detection. Now, as soon as you ask for the color of the detected object, for example, you can see that the bouding boxes drift from object, and the more details you ask for, the more they drift … really strange behaviour 