EXAONE Deep Unveiled — Setting a New Standard for Reasoning AI

LG AI Research has unveiled EXAONE Deep, a high-performance reasoning-specialized model that maximizes mathematical logic, scientific concept reasoning, and programming problem-solving ability in preparation for the Agentic AI era.

The EXAONE Deep 32B model showed overwhelming performance compared to competing models on high-difficulty math benchmarks. In particular, it recorded a high score on the math section of the 2025 CSAT (Korean SAT), scored 90.0 on AIME 2024, the US olympiad standard, and achieved results on par with DeepSeek-R1 (671B) on AIME 2025.

Key performance metrics are as follows:

• Math: The 32B model achieved a performance edge on high-difficulty benchmarks despite being just 5% the size of competing models
• Science & Coding: The 7.8B and 2.4B models both ranked 1st on major benchmarks (GPQA Diamond, LiveCodeBench)
• MMLU: The 32B model recorded a score of 83.0, proving the best performance among domestic (Korean) in-house models

Immediately after release, the EXAONE Deep 32B model was listed on the Notable AI Models list by Epoch AI, a US non-profit research organization, demonstrating global-level performance. In addition, the lightweight 7.8B and 2.4B models were designed to deliver excellent performance in on-device environments as well.