On March 18th, LG AI Research introduced EXAONE Deep via a blog post, positioning it as an innovative "Agentic AI" model. EXAONE Deep stands out with its capacity to autonomously propose and validate hypotheses, marking it as the first open-source reasoning AI model in South Korea. Boasting an impressive 32 billion parameters, this model achieved a remarkable score of 94.5 on the math section of the College Scholastic Ability Test (CSAT) in South Korea and a high score of 95.7 on the MATH-500 math test. Furthermore, LG AI Research introduced the EXAONE Deep-2.4B version, tailored for end devices. Despite having only 7.5% of the parameters of the original model, it retains 86% of its performance. This version supports local data processing without requiring external server connections, significantly bolstering security.
