LG's ChatEXAONE vs ChatGPT: Korean AI Model Performance Test Results
LG has quietly entered the AI chatbot race. After keeping its EXAONE technology under wraps for internal use, the South Korean tech giant finally opened its ChatEXAONE service to external users in July 2025.
But how does it actually perform against established players like ChatGPT and Google Gemini? IT Chosun put these AI models head-to-head in real-world scenarios to find out.
ChatEXAONE runs on EXAONE 3.5, which uses a "Mixture of Experts" architecture that combines smaller specialized models for better efficiency. Unlike consumer-focused competitors, LG targets enterprises and institutions, requiring corporate email verification for access.
The model comes with two standout features:
- "Deep" mode: Analyzes complex questions step-by-step for comprehensive answers
- "Dive" function: Lets users search specific sources like academic papers or YouTube
When asked for summer convenience store sales tips, ChatEXAONE delivered surprisingly detailed business insights. While ChatGPT and competitors offered standard advice (cold drinks, frozen foods, ice products), EXAONE broke down strategies into specific categories:
- Summer product enhancement with themed store decorations
- Weather-linked marketing campaigns
- Customer experience improvements like sunscreen and hat sales
- Local area targeting strategies
For ESG article planning, ChatEXAONE again showed superior business understanding, clearly separating Environmental, Social, and Governance elements while suggesting current news-based case studies.
EXAONE 3.5 achieved top performance among global open-source AI models across 20 benchmarks, including real-world usability and long text processing. In creative writing tests, it generated engaging science fiction with titles like "Memories of the Burning Sun," complete with structured chapters and epilogues that captured modern sensibilities.
Math Problems
ChatEXAONE struggled with probability questions from South Korea's college entrance exams, producing incorrect answers where ChatGPT and Google Gemini solved them correctly with clear explanations.
Information Currency
The model showed knowledge gaps about recent events. When asked about LG EXAONE's development team, it failed to mention that former LG AI Research chief Bae Kyung-hoon was appointed as South Korea's Minister of Science and ICT in July 2025, while ChatGPT included this updated information.
Sensitive Content Handling
ChatEXAONE gave mixed responses to politically sensitive questions. While it correctly identified territorial disputes, it showed inconsistent handling of potentially offensive content compared to competitors.
ChatEXAONE supports 32,000 tokens (processing 20,000 Korean words simultaneously) with plans to expand to 128,000 tokens in early 2025. The platform has obtained ISO certification for processing sensitive national technology documents, highlighting its enterprise security focus.
EXAONE models have exceeded 5 million downloads, a record for Korean AI companies, spanning three variants: 2.4B parameters (mobile devices), 7.8B parameters (general use), and 32B parameters (specialized applications).
ChatEXAONE isn't trying to be another ChatGPT. LG's strategy focuses on enterprise applications where detailed business analysis matters more than mathematical precision or encyclopedic knowledge.
For strategic planning and business intelligence, ChatEXAONE often outperformed established competitors. For math problems and current events, it lagged behind.
As LG noted, "ChatEXAONE isn't a general-purpose service like ChatGPT or Google Gemini, but a beta service targeting companies, institutions, and schools." The question is whether businesses will value specialized analytical capabilities over broader knowledge and computational accuracy.
With LG planning continuous improvements based on user feedback during its free beta period, ChatEXAONE represents an interesting alternative approach in the increasingly crowded AI landscape-one that prioritizes depth over breadth.