t5-small and t5-base are both translation models. See each entry for specifics.
T5-small is the 60M-parameter variant of Google's Text-to-Text Transfer Transformer, casting all NLP tasks as seq2seq problems. It was influential in establishing the unified text-to-text training paradigm but is outdated for production use.
T5-base is the 220M-parameter version of Google's T5, providing a better accuracy-speed trade-off than T5-small for seq2seq tasks while remaining significantly smaller than T5-large. It established the text-to-text format for multi-task NLP.
Pick based on your compute budget and specific task requirements.