LG releases open-weight AI model outperforming GPT-5 mini

Friday 10 April 2026 - 11:20

By: Dakir Madiha

LG releases open-weight AI model outperforming GPT-5 mini

LG AI Research introduced EXAONE 4.5 on Thursday, a vision-language model blending a proprietary vision encoder with a large language model to handle text and images together. The South Korean firm's research unit claimed superior average scores over OpenAI's GPT-5 mini and Anthropic's Claude Sonnet 4.5 across 13 metrics on AI vision processing and reasoning.

The model earned 77.3 points on STEM benchmarks in science, technology, engineering, and mathematics, which LG called world-class. It also beat Google's latest model on coding benchmarks and complex chart analysis, per Chosun Biz.

EXAONE 4.5 packs 33 billion parameters total, including 1.2 billion from the vision encoder, with a 262,144-token context length. Efficiency stands out: LG cut active parameters to 3.3 billion, one-seventh prior size, while matching text reasoning via hybrid attention architecture.

It supports six languages: Korean, English, Spanish, Japanese, and Vietnamese. LG released open weights on Hugging Face for research and academic use, its first such vision-language model.

LG targets industrial uses over chatbots. EXAONE 4.5 offers "real-world reasoning" to parse complex unstructured data like design drawings, financial statements, and tech contracts on factory floors.

The model fits LG's "K-EXAONE" project to grow into speech, video, and physical realms. "EXAONE 4.5 signals the multimodal era, fully grasping visuals beyond text," said Lee Jin-sik, EXAONE Lab head at LG AI Research. "We aim to extend to speech, video, and physical sites for AI that judges and acts in industry."