LG Group has developed a hybrid artificial intelligence (AI) that combines a large language model (LLM) with an inference AI model capable of independently formulating and verifying hypotheses.

LG AI Research unveiled South Korea’s first hybrid AI model, EXAONE 4.0, on Tuesday. The model comprising two models with 32 billion (32B) and 1.2 billion (1.2B) parameters, similar to synapses in the human brain.

The 32B model demonstrated expert-level capabilities, passing written exams for six national professional certifications including physician, dentist, Korean pharmacist, customs broker, appraiser, and claims adjuster, according to LG.

The 1.2B model is designed for practical deployment in devices such as home appliances, smartphones, automotive electronics, and robots. It processes data quickly and securely within the device itself without the need for external server connections, offering advantages in privacy and security.

EXAONE 4.0 also proved that it can compete against open-weight models from the United States, China, and France. The EXAONE 4.0 32B model scored 81.8 on the massive multitask language understanding (MMLU) pro test, 66.7 on LiveCodeBench v6 for coding skills, 75.4 on the Graduate-Level Google-Proof Q&A’s science problem-solving question set GPQA-Diamond, and 85.3 on AIME 2025 for math problem-solving.

These results outperform similar-sized competitors and are comparable to, or even top, much larger models such as DeepSeek’s R1 671B model and Alibaba’s Qwen3 235B model. In the Instruction Following Evaluation (IFEval) benchmark particularly, which measures an AI’s ability to understand and follow commands, EXAONE 4.0 scored 84.8, outperforming DeepSeek (81.2) and Qwen (83.2).

Only U.S-based Anthropic PBC and China’s Alibaba have released hybrid AI models to date, according to LG AI Research, while OpenAI has disclosed that it is developing GPT-5 as a hybrid AI.

[ⓒ 매일경제 & mk.co.kr, 무단 전재, 재배포 및 AI학습 이용 금지]