SEOUL, April 9, 2026 — LG AI Research has officially announced the release of its latest multimodal artificial intelligence model, EXAONE 4.5, marking a significant milestone in the global AI race. The new model, which simultaneously processes and reasons through both text and visual data, has reportedly outperformed leading models from OpenAI, Google, and Anthropic in several key performance benchmarks.
A Breakthrough in Vision-Language Integration
EXAONE 4.5 is a Vision Language Model (VLM) that seamlessly integrates a proprietary vision encoder with a Large Language Model (LLM). Unlike conventional models that often struggle with complex layouts, EXAONE 4.5 excels at interpreting "industrial-grade" documents. This includes intricate contracts, technical blueprints, financial statements, and scanned documents, where the AI must not only read text but also understand the spatial relationship of charts, tables, and diagrams.
Dominating the Benchmarks
The performance metrics released by LG AI Research suggest a shift in the AI hierarchy. In the STEM (Science, Technology, Engineering, and Mathematics) category, EXAONE 4.5 achieved an average score of 77.3. This figure places it ahead of OpenAI’s GPT-5 mini (73.5), Anthropic’s Claude 4.5 Sonnet (74.6), and Alibaba’s Qwen 3 235B (77.0).
Furthermore, the model showcased exceptional prowess in specialized reasoning. On the LiveCodeBench v6, a premier indicator of coding capability, EXAONE 4.5 scored 81.4, edging out Google’s newest Gemma 4 (80.0). In the ChartQA Pro benchmark, which evaluates the ability to analyze and infer information from complex charts, the model scored 62.2, solidifying its position as a global leader in document intelligence.
“Achieving high scores in visual evaluation indicates that our AI has moved beyond simple OCR (Optical Character Recognition) or data identification,” said a representative from LG AI Research. “It means the model possesses the cognitive depth to understand context and provide sophisticated answers to complex queries.”
Efficiency Through Innovation
One of the most striking aspects of EXAONE 4.5 is its efficiency. With 33 billion parameters (33B), it is approximately one-seventh the size of the previous 'K-EXAONE' model released late last year. Despite the smaller size, it delivers equivalent or superior performance in text reasoning. This was made possible by LG's proprietary Hybrid Attention architecture and Multi-token Prediction-based high-speed inference technology, which allows the model to process information faster and more accurately with fewer computational resources.
Toward a Physical AI Future
The launch of EXAONE 4.5 is seen as a strategic move by LG to dominate the industrial AI sector, where precise document analysis is critical. Lee Jin-sik, Head of the EXAONE Lab at LG AI Research, emphasized the significance of this transition.
“EXAONE 4.5 represents LG AI’s successful entry into the multimodal era, where the AI understands visual information as fluently as text,” Lee stated. “Starting with this model, we intend to expand the AI's scope to include audio, video, and even physical environments. Our ultimate goal is to create an AI that can practically judge and act within real-world industrial settings.”
As the competition between Silicon Valley and Asian tech giants intensifies, LG’s latest breakthrough suggests that the future of AI may not just belong to the largest models, but to the smartest and most efficient ones.
[Copyright (c) Global Economic Times. All Rights Reserved.]
