Qwen2.5-VL-7B vs 32B

Has anyone tested Qwen2.5-VL-7B vs 32B specifically on identifying checkmarks/signatures on photocopied scanned forms (1980s-2010s government forms, low-quality scans)? 7B reliably transcribes form labels but cannot distinguish selected from unselected checkboxes on degraded scans in my testing. Looking for empirical comparison before investing in the larger model.