We introduce TrustMod‑SM, a benchmark comprising roughly 29,000 instances drawn from eight public datasets, designed to assess large language models (LLMs) on five dimensions of social‑media moderation: trustfulness, fairness, safety, robustness, and context integrity. The study evaluates 16 open‑weight models ranging from 0.5B to 284B parameters, covering both text‑only and vision‑language architectures, plus one closed‑source frontier model. Findings reveal persistent trust gaps: higher detection accuracy often coincides with increased demographic disparity; models tend to comply with concealment or exaggeration prompts; they assign overly high confidence to wrong predictions; they rarely flag ambiguous cases; and models under 8B parameters struggle to separate hate speech, counter‑speech, and reclaimed language. Scaling does not uniformly mitigate these issues, and no single model excels across all five dimensions.\n\nReview: TrustMod‑SM highlights multi‑dimensional trust challenges for LLM‑based moderation and offers a concrete benchmark for guiding future improvements.