LG’s hybrid AI ‘Exa1 4.0’ Achieves Breakthroughs, Outperforming Global Competitors
LG Group has announced a notable advancement in artificial intelligence wiht the prosperous progress of its first hybrid AI model, ExaOne 4.0. This groundbreaking AI demonstrates the capability to pass professional certification exams, including those for doctors and oriental medicine practitioners. ExaOne 4.0 uniquely combines a large language model (LLM), excelling in natural language understanding and generation, with a “reasonable AI model” designed for hypothesis generation and self-verification. LG AI researchers position ExaOne 4.0 among world-leading hybrid AIs, comparable to those developed by US-based Anslo Pic and China’s Alibaba.
ExaOne 4.0 is offered in two parameter sizes: 32 billion (32B) and 1.2 billion (1.2B). The 32B model has garnered high praise for its professional knowledge and problem-solving abilities,evidenced by its success in passing six national certification writen tests,including those for doctors,dentists,herbal medicine practitioners,customs evaluators,appraisers,and damages assessors. The 1.2B model offers practical applications for a wide range of devices such as home appliances, smartphones, automotive electronics, and robots. Its ability to process information quickly and securely within the device, without external server connection, enhances personal information protection and overall security.
performance benchmarks reveal ExaOne 4.0 is competitive with leading “open weight” models from the US, China, and france. The ExaOne 4.0 32B model achieved an remarkable 81.8 points in the MMLU Pro benchmark for AI knowledge and problem-solving, 66.7 points in Livecodebench V6 for coding ability, 75.4 points in GPQA-DIAMOND, and 85.3 points in AIME 2025. These scores are on par with, and in some cases surpass, AI programs of similar or even larger capacities, such as Deepseek R1 (671B) and Alibaba’s Qwen3 (235B). Notably, ExaOne 4.0 scored 84.8 points in the IFEVAL benchmark, which assesses command understanding and execution, outperforming Deepseek (81.2 points) and Qwen3 (83.2 points).
[By Kim Dong-eun, Reporter]
If you liked this article, Please press like.
great 0
