LG Unveils EXAONE Deep: Korea's First Open-Source Reasoning AI

San Jose, CA / Seoul, South Korea - LG AI Research has released "EXAONE Deep," a reasoning artificial intelligence model, as open source. The announcement coincides with NVIDIA's annual San Jose, California, developer conference.

The model represents what LG terms "Agentic AI," which can formulate and verify hypotheses independently. 

According to the company, EXAONE Deep is the first Korean-developed reasoning AI to compete in the global market dominated by a few companies with foundation models.

The primary version, EXAONE Deep-32B with 32 billion parameters, operates at approximately 5% of the size of larger competitor models while achieving comparable performance. 

In benchmark testing, the model scored 94.5 in the 2025 CSAT mathematics section and 95.7 points in MATH-500, an index measuring mathematical problem-solving capabilities.

EXAONE Deep-32B recorded 66.1 points for scientific reasoning in the GPQA diamond test, which evaluates doctoral-level scientific problem-solving across physics, chemistry, and biology. In coding assessments, it achieved 59.5 points in LiveCodeBench.

The company has simultaneously released two smaller versions: EXAONE Deep-7.8B, which maintains 95% of the performance at 24% of the size, and EXAONE Deep-2.4B, an on-device model retaining 86% of the performance at 7.5% of the original size.

The on-device model processes data locally without external server connections, potentially offering enhanced security. 

LG AI Research indicates plans to expand on-device AI applications across smartphones, automobiles, and robotics through collaborations with LG Electronics and LG U+.

The US non-profit AI research organization Epoch AI has included EXAONE Deep in its "Notable AI Models" list.