Protein function prediction is critical for a wide range of applications in biology, spanning from functional genomics to protein design and genome evolution, among others. However, accurately predicting protein function remains a longstanding challenge in computational biology, especially for non-model organisms. Traditional methods based on sequence similarity often fail to annotate a significant proportion of proteins. The emergence of protein language models has significantly improved this process, enabling more accurate and comprehensive functional annotation. In this work, we highlight how the ProtTrans language model outperforms other tools in per-protein annotation, offering a more precise approach to predicting protein function. We also introduce functional annotation based on embedding space similarity (FANTASIA; available at https://github.com/MetazoaPhylogenomicsLab/FANTASIA ), a tool developed to harness these advances for large-scale annotation of uncharacterized proteomes. We provide a detailed overview of how to use FANTASIA, interpret its outputs, and demonstrate its utility in three case studies: (a) enrichment analyses from transcriptomics data, (b) assigning novel functions to unannotated genes in model organisms, and (c) identifying genes involved in important functions in non-model organisms. These results demonstrate the potential of protein language models to advance functional annotation in diverse biological contexts.