EXAONE 4.5

What it is

EXAONE 4.5 is a vision language model — a model that takes images as well as text as input — from LG AI Research. It is the first open-weight model of that kind from the lab: the trained weights are published so that others can download and run them.^[1][2] The model has 33 billion parameters in total, 1.2 billion of which belong to the vision encoder, the part that turns an image into representations the language side can work with.^[2] A technical report describing the model appeared on 9 April 2026.^[1]

How it was built

LG did not start from a blank language model. EXAONE 4.5 was built by adding a dedicated visual encoder to the existing EXAONE 4.0 framework.^[1] Pretraining was natively multimodal: images and text were trained on together, rather than a text-only model being adapted to images afterwards. The training mixture was weighted toward document corpora, so a large share of the visual material the model learned from was pages of documents rather than everyday photographs.^[1]

Long context and languages

The context window was extended to 256K tokens, a length aimed at the long documents that turn up in enterprise work.^[1] The model card lists support for English, Korean, Spanish, German, Japanese and Vietnamese.^[2]

What LG claims

LG states that the model beats same-size state-of-the-art models on document understanding and on Korean contextual reasoning.^[2] These are the lab's own comparisons at the 33-billion-parameter size. The two areas they single out line up with visible choices in the design: the document-heavy training data and the extended context window suit long pages of text, and Korean is one of the languages the model card lists.^[1][2]

Release and licence

The 33B weights were uploaded to Hugging Face on 4 April 2026, a few days ahead of the technical report.^[3][1] The model keeps the EXAONE licence rather than an OSI-approved one.^[2] That distinction is worth knowing. "Open weight" means the parameters are downloadable; an OSI-approved licence would additionally satisfy the usual definition of open source, in which the terms of use are approved by an open-source body. Here users work under LG's own terms.

Where it sits

EXAONE 4.5 carries LG's EXAONE line from text into images. The emphasis on documents, long context and enterprise-scale material points the model at contracts, reports and scanned pages rather than at casual photo conversation. The open weights matter for the same reason: an organisation can run the model inside its own infrastructure instead of sending material to an outside service, which is often the deciding question when the input is internal paperwork.

For a reader trying to place the model, three details are the most informative. The parameter count puts it in the class of models large enough to be useful but small enough to serve without a cluster of accelerators. The vision encoder is a modest fraction of the total, which is consistent with a design in which most of the capacity went into language and document reasoning rather than into image perception.^[2] And the native multimodal pretraining means the model never had a text-only phase that images were bolted onto later.^[1]

As with any model card, the performance statements describe what the developer measured against its chosen comparison set; they are not a neutral scoreboard. What can be said without qualification is the shape of the thing: a 33-billion-parameter model, 1.2 billion of them visual, trained on images and text with a document-heavy diet, reading up to 256K tokens of context, covering six languages, published as downloadable weights under a custom licence.^[1][2][3]