LG AI Research

What it is

LG AI Research is the organisation behind EXAONE, a family of large language models that it documents in technical reports on arXiv and publishes under the LGAI-EXAONE organisation on Hugging Face ^[1][5]. The series is the clearest public view of what the lab builds: instruction-tuned general models, dedicated reasoning models, and a vision language model ^[1][4][5].

How the series grew

EXAONE 3.0 was published on 7 August 2024 and was the first EXAONE model released openly ^[1]. That first release was narrow by design: of several sizes built internally, only the 7.8B instruction-tuned version was made public ^[1].

EXAONE 3.5 followed on 6 December 2024 in three instruction-tuned sizes, 2.4B, 7.8B and 32B, all with a 32K context window ^[2]. The technical report records the shape of the middle model: 6.98 billion parameters excluding embeddings, 32 layers and a 102,400-token vocabulary ^[2].

The EXAONE 4.0 report was posted on 15 July 2025 with two sizes, a 32B model and a 1.2B on-device model ^[3]. This generation combined two earlier lines, the usability of 3.5 and the reasoning ability of EXAONE Deep, and added agentic tool use, meaning the model can call tools as part of carrying out a task ^[3]. Multilingual support widened from English and Korean to Spanish ^[3].

EXAONE 4.5 appeared on 9 April 2026 as the company's first open-weight vision language model, in a single 33B size paired with a 1.2B vision encoder ^[4]. Context was extended to 256K tokens for enterprise long documents, and German, Japanese and Vietnamese were added ^[4].

A distinctive release policy

Every EXAONE release keeps a custom EXAONE licence rather than an OSI-approved open-source licence ^[2][5]. The weights are published openly, but the terms of use are set by the lab itself rather than by a general open-source licence. The family was also extended with EXAONE Deep reasoning models in March 2025 ^[2][5].

Across the releases the pattern is a spread of sizes rather than one flagship: a small on-device model at one end, a 32B-class model at the other, and a mid-size model documented in the most detail ^[1][2][3][4].

Where it sits

The EXAONE line shows a research lab moving from a single, narrowly released model towards a family that covers several deployment settings at once: on-device, mid-size and large; text and then vision; English and Korean first, then Spanish, German, Japanese and Vietnamese ^[1][3][4]. Because each step comes with a technical report giving parameter counts, layer counts and vocabulary sizes, the series also works as a public record of how one organisation's model line changed from one generation to the next ^[2][3][4]. For readers outside the lab, that combination of open weights plus detailed reports is what makes the organisation legible: the licensing stays proprietary, while the architecture and training choices are described in the open ^[2][5].