LG Released the Latest LLM, but the No. 1 Download Was Seoul National University's 4-Year-Old Small Model

Although South Korean AI companies have successively released large language models (LLMs), the number one downloaded open model directly downloaded and used by developers was not LG's latest LLM, but a financial text classification model released by a Seoul National University lab in 2022.

LG's EXAONE, Upstage's SOLAR, and SK Telecom's A.X. South Korean AI companies are each competing by releasing large language models. A significant portion of this competition takes place in each company's commercial APIs and services. However, there is a separate demand for developers to directly download model files and run them on their own servers, and this demand is counted by the number of downloads on the AI model sharing platform HuggingFace. As a result of News Epoch aggregating the downloads of South Korean AI organizations and researcher accounts over the last 30 days, the demand in this area differed greatly from the announced competitive landscape.

No. 1 Download is a 4-Year-Old Financial Classification Model

Based on the last 30 days, the total downloads of the tracked accounts amounted to 2.63 million. Among them, the single most downloaded model was the financial text classification model KR-FinBert, released by the Seoul National University Natural Language Processing Lab in 2022, recording 256,000 downloads and surpassing the latest model EXAONE-4.5-33B (247,000 downloads) released by LG this year.

Source: HuggingFace Hub API (Collected on July 1, 2026), last 30 days downloads.

Excluding EXAONE, the top ranks are occupied by small task models released in 2022, such as the document OCR model Donut, the language model KcBERT, and the question answering model KoELECTRA. Because these models are directly inserted into search, classification, and document processing pipelines rather than being called via an API, their download numbers are close to actual usage. This means that even after large LLMs were successively released, the models downloaded for practical pipelines remain in the previous generation.

Generative LLM Downloads Were Concentrated on EXAONE

Looking at it by account, first place was LG AI Research's EXAONE account, recording 885,000 downloads and accounting for 33.6% of the total. Generative LLM downloads are mostly EXAONE, and excluding LG, the combined generative models of the remaining accounts fall short of 1% of the total. The representative model of Upstage's SOLAR had 64,000 downloads, and the representative model of SK Telecom's A.X had 31,000 downloads, which are one-fourth and one-eighth of the EXAONE representative model, respectively.

However, since SOLAR and A.X are models that each company also provides through commercial APIs, it cannot be concluded that overall usage is low just because downloads are low. What the download count shows is the demand for directly downloading and operating models, and within this demand, generative LLMs are concentrated on EXAONE.

Open Models Directly Downloaded and Used Are a Minority

In the area where developers directly download and use them, the demand for generative LLMs is concentrated entirely on EXAONE, and the number one download is not the latest LLM but a 4-year-old small task model.

There is a limit to download counts. Small model downloads include a portion of repeated calls by libraries and automated pipelines, so it is impossible to know exactly which model the user consciously chose. However, even considering this, the gap between EXAONE and the remaining generative models, and the situation where the latest LLMs are pushed aside by a 4-year-old small model, are not overturned. There are several companies that have released new LLMs, but the open models actually downloaded and used by developers are a minority.

This article was written based on OpenIndex, operated by News Epoch to quantitatively track the South Korean AI open-source ecosystem. The data is the last 30 days of downloads collected on July 1, 2026, via the HuggingFace Hub API, and does not include usage through commercial APIs or services.