LG Releases EXAONE 4.5 — The 33B Open VLM That Outscores GPT-5-mini and Claude 4.5 Sonnet on STEM

LG AI Research has released EXAONE 4.5, a 33B-parameter multimodal Vision-Language Model that claims top performance among open models on scientific and technical reasoning tasks. The model scores an average of 77.3 across five STEM benchmarks — beating GPT-5-mini at 73.5 and Claude 4.5 Sonnet at 74.6.

The model is fully multimodal, handling both text and images, with expanded language support spanning Korean, English, Spanish, German, Japanese, and Vietnamese. LG has made the weights available on HuggingFace under a research and academic use license, with commercial licensing available via direct inquiry.

EXAONE 4.5 follows LG's EXAONE 4.0 series, which focused primarily on Korean and English language tasks. The expansion to multilingual STEM reasoning represents a meaningful capability leap and positions the model for use in scientific research tooling, technical documentation systems, and multilingual educational applications.

The context here matters: LG is not traditionally a frontier AI company. That a consumer electronics and industrial conglomerate is shipping models that outperform Anthropic's current-generation Sonnet on STEM benchmarks says something important about how broadly AI capability is diffusing across the industry. Enterprise buyers who assumed OpenAI and Anthropic were the only options for serious technical applications should be paying attention.

For builders, the research-use license is a constraint worth noting — you'll need to negotiate commercial terms for production applications. But for academic researchers, the model is available today.