This paper explores the potential of Large Language Models (LLMs) to semantically enrich dataset metadata, comparing proprietary and open-source models. Employing a five-stage workflow that includes data loading, preprocessing, profiling, enrichment with LLMs, and evaluation, 16 real datasets from the Ecuador Open Data Portal were processed. For profiling, statistical properties and representative examples were extracted for each column using Python and converted into structured JSON prompts. The enriched descriptions were then evaluated using Cosine Similarity and Recall-Oriented Understudy for Gisting Evaluation - Longest common subsequence (ROUGE-L) F1 against reference descriptions from the official data dictionaries. All models exceeded 0.6 in semantic similarity, with Llama 4 Maverick being particularly prominent. In addition, open-source models demonstrated competitive performance when compared to their proprietary counterparts, particularly in domains with standardized terminology. These findings validate the feasibility of LLMs for automating metadata enrichment tasks, offering significant time and resource savings. Moreover, they emphasize the practical benefits of open-source solutions in contexts where data privacy, transparency, and licensing flexibility are crucial.