EXAONE 4.5: LG AI Research Ships 33B Multimodal VLM with Hybrid Attention and 6-Language Support

Hugging Face·medium signalXBlueskyLinkedInCopy link

LG AI Research released EXAONE 4.5 on April 9 — a 33B vision-language model (1.2B vision encoder + LLM) using a proprietary hybrid attention architecture. Claims to beat OpenAI and Google models on specific benchmarks at one-seventh the parameter count of their previous K-EXAONE model. Supports Korean, English, Spanish, German, Japanese, and Vietnamese. Open weights on Hugging Face for research. A strong mid-size multimodal option for builders needing multilingual VLM capabilities without frontier-scale compute.