Answers are grounded in this paper and cite their sources.
This study evaluates the performance of GEMMA3, a lightweight large language model, in solving SAT Math problems within constrained computational environments. Two model variants-GEMMA3 1B and 4Bwere tested using 100 standardized SAT Math questions. While the 4B model achieved higher accuracy (83 %), it incurred significantly greater inference time and memory usage. In contrast, GEMMA3 1B offered faster response times and lower RAM consumption, with slightly higher CPU load. The results underscore the trade-off between accuracy and efficiency, offering insights for selecting AI tools suited to resource-limited educational contexts.
No takes yet. Share an insight, caveat, or question.
Lertyosbordin et al. (2025) studied this question.