EXAONE 4.5 33B is a multimodal vision-language model developed by LG AI Research. It represents the first open-weight multimodal release in the EXAONE series, natively integrating visual comprehension with large-scale language processing. The model is specifically engineered for industrial and enterprise intelligence, demonstrating high proficiency in science, technology, engineering, and mathematics (STEM) fields, as well as complex document understanding across six languages, including Korean, English, and Japanese.
The architecture features a 33 billion parameter dense structure, comprising a 31.7 billion parameter language backbone and a proprietary 1.2 billion parameter vision encoder. It employs a Hybrid Attention pattern that alternates between sliding-window and global attention to manage its expansive 262,144-token context window efficiently. This design allows the model to process high-resolution images and long-form documents simultaneously, making it suitable for analyzing technical blueprints, financial reports, and industrial diagrams.
While the model includes an internal reasoning mode (Thinking mode) by default, the Non-reasoning mode is utilized to minimize latency in general-purpose tasks. In this mode, the model bypasses the generation of intermediate thought tokens to deliver direct responses. This configuration maintains the model's core knowledge and document-parsing accuracy while significantly improving inference speed and reducing token consumption for straightforward queries.
For optimal performance in non-reasoning tasks, developers recommend specific sampling configurations. For general text and multimodal inputs, a temperature of 1.0 and a top-p of 0.95 are suggested. For specialized tasks such as Optical Character Recognition (OCR) or Korean-centric queries, a lower temperature of 0.6 and a top-k of 20 are recommended to enhance response consistency and structural accuracy.