LG AI Research
Open-weight 33B vision-language model from LG AI Research for text-image reasoning, document understanding, STEM tasks, and Korean contextual reasoning.
EXAONE 4.5 is a credible 33B open-weight VLM with a differentiated emphasis on document understanding, text-image reasoning, STEM, and Korean visual context. Its official technical report claims an 81.4 LiveCodeBench v6 result for the 33B Reasoning variant, but the current independent LiveCodeBench leaderboard does not list 4.5; this supports a solid, not top-tier, coding score. The official material substantiates multimodal availability, while public evidence is insufficient to place it near Claude Opus 4.8 or GPT-5.5 on broad capability. Relative to open-weight GLM-4.6, EXAONE scores higher in multimodal I/O because it is explicitly a released VLM with document-oriented positioning; its technical and reasoning scores are also modestly higher given the reported LiveCodeBench result. It remains below Gemini 3.5 Flash on multimodal ecosystem maturity, serving availability, and developer integration. The open weights and Artificial Analysis’ $0 estimated token pricing support strong cost effectiveness for operators able to self-host, but that estimate is not a provider API price. Evidence quality is moderate rather than high: central performance evidence is vendor-reported, and no DeepSWE/DataCurve, SWE-bench, Terminal-Bench, Aider, LMArena, or Arena-Hard result was supplied. LG’s official page and GitHub repository improve reproducibility and developer confidence, but supplied sources do not specify license terms, hosted-service pricing, measured inference speed, or broad adoption. Scores therefore favor the documented VLM strengths while discounting missing independent agentic and preference evidence.
