EXAONE 4.5 33B is an open-weight vision-language model developed by LG AI Research. Released in April 2026, it represents the first multimodal entry in the EXAONE series to be made available for public research and educational use. The model is built on a 33-billion parameter architecture, comprising a 31.7B parameter language model integrated with a proprietary 1.29B parameter vision encoder. It is designed for high-performance reasoning across both text and visual inputs, supporting six languages: Korean, English, Japanese, Spanish, German, and Vietnamese.
The model utilizes a unique Hybrid Attention architecture that alternates between sliding window attention and global attention to optimize processing efficiency. It also incorporates a Multi-Token Prediction (MTP) mechanism, which significantly enhances inference throughput. For visual processing, the dedicated vision encoder employs Grouped Query Attention (GQA) and 2D Rotary Positional Embeddings (RoPE) to accurately interpret spatial data, making it particularly effective for document understanding, technical drawings, and complex charts.
EXAONE 4.5 33B is highly optimized for industrial and academic applications, demonstrating specialized performance in STEM (Science, Technology, Engineering, and Mathematics) benchmarks and document-centric reasoning tasks. It supports a large context window of 262,144 tokens (256K), allowing it to process extensive documents and long-range visual-textual contexts. The model was trained with a native multimodal pre-training approach, ensuring that visual and linguistic features are fused from the early stages of learning rather than being combined post-hoc.
Performance and Prompting
Benchmarks show that the model excels in complex visual reasoning, outperforming several contemporary mid-tier commercial models in STEM evaluations and chart analysis. Notably, it features an integrated reasoning mode enabled by default, which allows the model to engage in internal chain-of-thought processing before delivering a final answer.
For optimal results, users are encouraged to use specific generation parameters based on the task. For OCR and document-related tasks, a temperature of 0.6 and top_p of 0.95 are recommended. For general text-only inputs, a higher temperature of 1.0 is suggested. The model is also tuned to provide structured answers using a \boxed{} format for mathematical and precise reasoning queries to ensure parsing accuracy.