LG Unveils 'EXAONE Deep' — 'The Era of Reasoning AI'

March 18, 2025: LG AI Research officially unveiled high-performance reasoning-specialized AI model "EXAONE Deep." This model is a Reasoning-specialized AI model with mathematical thinking, scientific concept understanding, and programming problem-solving capability -- containing technological evolution targeting the "Agentic AI" era gaining global attention. EXAONE Deep immediately after its public release was selected by US AI research organization Epoch AI as a "Notable AI Model" -- evaluated as a reasoning AI model standing shoulder to shoulder with OpenAI o3-mini, DeepSeek R1, and other global leading models in its category. Three model sizes: 2.4B (lightweight, for edge deployment), 7.8B (balanced performance/efficiency), 32B (maximum performance for enterprise). Performance benchmarks: on MATH-500 (mathematics reasoning), AIME 2024 (advanced mathematics competition), and HumanEval (coding) -- the 32B model achieves scores comparable to DeepSeek R1 and OpenAI o3-mini; particularly strong on Korean-language mathematical reasoning where training data includes Korean textbooks and exam problems. The Agentic AI positioning: reasoning models that can break complex problems into steps and verify intermediate results are essential for agentic applications (AI that takes multi-step actions to complete goals); EXAONE Deep reasoning architecture enables autonomous task completion rather than single-turn question answering; LG is positioning EXAONE Deep as the reasoning backbone for enterprise AI agents across LG group companies (manufacturing, home appliances, energy) and external enterprise customers. Open source release: EXAONE Deep 2.4B and 7.8B released on Hugging Face under EXAONE Community License (free for research and commercial use under 100M USD annual revenue threshold).