Large Language Models (LLMs) have achieved remarkable progress in natural language understanding and generation, yet their evaluation remains heavily centered on English and Western cultural contexts, limiting their reliability in culturally diverse and low-resource settings. We introduce ParsBench, a comprehensive benchmark for evaluating fine-grained culturally grounded commonsense reasoning in Persian. ParsBench consists of three complementary components: Expert-QA, Multimodal-QA, and Long-Form QA, collectively covering structured reasoning, multimodal cultural understanding, and open-ended generation across diverse cultural categories. We evaluate a broad range of multilingual and multimodal LLMs and observe substantial variation across models and cultural categories. Our results reveal persistent weaknesses in culturally grounded non-English reasoning and significant disparities between textual and multimodal understanding. Overall, the findings demonstrate that existing English-centric evaluations fail to capture important dimensions of cultural competence and highlight the need for culturally grounded benchmarks for more inclusive evaluation. Our code and dataset are available at https://github.com/ParsBench-CCS/ParsBench