EXAONE‑4.5‑33B release

Artificial intelligence models are prediction engines: they guess the next word or token, and newer ones can do that across images too. LG AI Research has now released EXAONE‑4.5‑33B, a 33‑billion‑parameter open‑weight model on Hugging Face. (huggingface.co) The release went live in April 2026 through LG AI Research’s Hugging Face account, with model files, a README, and separate FP8 and GGUF variants for different deployment setups. The Hugging Face collection describes EXAONE 4.5 as LG’s first open‑weight vision‑language model for “industrial intelligence.” (huggingface.co) A vision‑language model combines a language model with an image encoder, so the system can read text, inspect charts, and answer questions about documents or photos in one workflow. LG’s paper says EXAONE‑4.5 adds a 1.2‑billion‑parameter vision encoder to the EXAONE 4.0 32B base model. (arxiv.org) LG AI Research submitted the technical report to arXiv on April 9, 2026, under identifier 2604.08644. The paper says the model was built for document understanding, Korean contextual reasoning, and other industrial tasks where text and images appear together. (arxiv.org) That focus puts EXAONE‑4.5‑33B in a crowded part of the market: models large enough to handle complex work, but smaller than the biggest frontier systems that are expensive to run. LG’s model card lists 33 billion total parameters, while the collection page also offers compressed formats aimed at lower‑cost inference. (huggingface.co) The “open‑weight” label matters because developers can download the model weights and run them on their own infrastructure, subject to the model license. LG distributes EXAONE‑4.5‑33B under its EXAONE license, not a standard permissive software license. (huggingface.co) LG is framing the model around factories, manuals, forms, and other enterprise documents rather than consumer chatbots. The technical report says multimodal training was designed to improve performance on visually rich industrial material, including document intelligence workloads. (arxiv.org) This is not LG’s first EXAONE release, but it is a step beyond the earlier text‑first open models. LG AI Research’s 2024 EXAONE 3.0 paper described a 7.8‑billion‑parameter open model, while the new 4.5 report presents a larger system that can natively process images alongside text. (arxiv.org) The immediate test is not whether EXAONE‑4.5‑33B is the biggest model available, but whether companies can actually deploy it for document pipelines and software agents. LG has published the weights, the paper, and multiple inference formats; the next evidence will come from adoption outside LG’s own stack. (huggingface.co)