LG AI Research Unveils 'EXAONE Deep' with Maximized Reasoning Performance

In preparation for the Agentic AI era, in which AI sets and verifies its own hypotheses, LG AI Research has introduced EXAONE Deep, a model with advanced reasoning capabilities. This model is a high-performance reasoning model specialized in mathematical logic, scientific concept understanding, and programming problem-solving.

EXAONE Deep demonstrated outstanding results on major benchmarks.

• Math: Despite being about 5% the size of competing models, the EXAONE Deep 32B model showed overwhelming performance on high-difficulty math benchmarks. It notably achieved the top score on the math section of the 2025 CSAT (College Scholastic Ability Test), and also earned a high score on AIME 2024, the US mathematics competition.
• Science & Coding: The 7.8B and 2.4B models ranked first on major benchmarks, proving their dominant performance.
• MMLU: The 32B model recorded a score of 83.0, achieving the best performance among domestic (Korean) models.

In particular, EXAONE Deep 32B showed performance on par with the DeepSeek-R1 (671B) model on AIME 2025, demonstrating outstanding training efficiency and cost efficiency relative to its model size. Immediately after release, it was listed in Epoch AI's 'Notable AI Models,' earning global recognition for its performance.