Large language models (LLMs) are increasingly used to generate psycholinguistic norms, yet their ability to approximate human sensorimotor ratings across languages and beyond central tendency remains poorly understood. The present study evaluated whether GPT-4.1 Nano and DeepSeek V3.2 approximate human sensorimotor norms across five languages (English, Italian, Chinese, Croatian, and Mexican Spanish) using demographically specified synthetic participants matched to each human norming sample. We assessed mean-level alignment, representational similarity, variability structure, model- and language-related differences, and the coherence of model-generated rationales for assigned ratings using an automated LLM-as-a-judge procedure. DeepSeek showed stronger rank-order and geometric alignment with human norms than GPT across languages, although distributional fit based on Jensen-Shannon divergence was more nuanced. Both models reproduced aspects of Croatian and Mexican Spanish norms, which are not publicly available, a pattern consistent with cross-linguistic generalization rather than simple retrieval of widely available norm sets. Perceptual dimensions showed stronger alignment than motor-effector dimensions, corroborating recent findings in the literature. Rationale analyses showed that explanations were generally coherent, but non-valid rationales were concentrated in motor-effector dimensions and often reflected cross-dimensional or incidental reasoning. Despite mean-level alignment, both models compressed human rating variability, indicating better approximation of population-level averages than individual disagreement. Mixed-effects analyses confirmed significant effects of model family and language, with the DeepSeek advantage in rank-order alignment broadly consistent across linguistic contexts. Sensorimotor norms generated across languages correlated most strongly for DeepSeek, followed by GPT and then human norms, suggesting that DeepSeek-generated profiles may be more linguistically uniform than human ratings. Overall, LLM-generated psycholinguistic norms can provide useful approximations of population-level semantic structure, but should not be treated as interchangeable with human norms: mean-level agreement does not necessarily imply comparable variability, distributional profiles, demographic fidelity, cross-cultural structure, or rationale consistency.