LG AI Research open-weights its first vision language model, EXAONE 4.5

Background

EXAONE 4.5 is the next step in LG AI Research's EXAONE line of models. The version it follows, EXAONE 4.0, worked with a context window of 131K tokens — the amount of text a model can hold in view at once ^[1]. EXAONE 4.5 keeps the family name but changes two things at the same time: it adds image input, and it is published with open weights.

What was released

LG AI Research released EXAONE 4.5 on 9 April 2026 ^[1]. It is the company's first multimodal model to be released with open weights ^[1][2]. Open weights means the trained parameters themselves are published, so people outside the company can download the model and run it on their own hardware rather than only calling it through an interface.

The weights were already on Hugging Face on 4 April 2026, five days before the technical report appeared ^[2]. That ordering — model card and weights first, paper afterwards — has become common for open releases, because it lets users start working with the model while the full write-up is still being finished.

What is in the model

Only one size was released: 33B parameters ^[2]. Within that total, 1.2B parameters belong to the vision encoder, the part that turns images into representations the language model can read ^[2]. The rest of the parameters carry the language side.

Context length grew from 131K in EXAONE 4.0 to 256K ^[1]. A longer context window means more material — a longer report, a bigger stack of pages — can be supplied in a single pass instead of being chopped into pieces and stitched back together.

Document understanding as the target

LG AI Research named document understanding as the main application it was aiming at, a target that lines up with LG's business domains ^[1]. Document understanding is the task of taking a page as it actually appears — text, tables, layout, form fields, stamps, scanned or photographed pages — and producing something a program can act on, such as extracted fields, a structured table, or an answer to a question about the document. It is harder than plain text reading because the meaning often lives in the arrangement: which number belongs to which column, which label sits next to which box. A vision encoder paired with a language model is one way to attack that, since the model can look at the page rather than at text recovered by a separate step.

Choosing this application also says something about who the model is for. Document processing is the kind of work that appears across large organisations — contracts, invoices, reports, forms — and a model built around it is aimed less at general chat and more at a defined, repetitive task.

The licence

EXAONE 4.5 keeps the licence used for earlier EXAONE releases rather than moving to an open-source licence ^[2]. So although the weights are open, the terms under which they may be used come from that document and not from a standard open-source licence. For anyone planning to build on the model, reading those terms is part of evaluating the release, in the same way the parameter count or the context length is.

What it signals

Two things stand out. First, a corporate research lab with a defined industry focus put a multimodal model out with open weights, which puts it in front of outside researchers and developers who can examine it, test it and report on what it does. Second, the model was designed around one use — reading documents — rather than around breadth, and its two headline specifications, the 256K context and the separate vision encoder, both serve that use. Whether the approach holds up in practice is something the people running the weights can now check for themselves.