First Tracking of 9 Korean AI Companies Reveals Gap Between Code Interest and Actual Model Usage

• First tracking of 9 Korean AI companies on GitHub and HuggingFace... Gap between code publication and actual model usage reaches up to 254 times

• LG ranks 1st with 930,000 model downloads despite lowest code interest... NCSoft at 3,676 downloads

• Naver CLOVA and Kakao Brain halt activities... Forging their own paths through reorganization and OpenAI partnership

AI companies release their proprietary technologies to the public for free. It is not an obligation, but a strategy. Just as Google dominated the smartphone ecosystem by releasing Android for free, releasing an AI model creates a virtuous cycle where developers build services on top of it, talent flocks to it, and external developers improve its performance. This is the core of the open-source strategy.

To see if this strategy actually works, one must look at two platforms: GitHub, where code is uploaded, and HuggingFace, where AI models are uploaded. The 'stars' clicked by developers on GitHub show interest in the code, while the number of downloads on HuggingFace shows the actual usage of the model. Overseas, there are regular analyses such as official HuggingFace reports or the GitHub Octoverse, but there has been no report cross-analyzing Korean AI companies across these two platforms. News Epoch has compiled for the first time the GitHub stars and HuggingFace downloads of 9 Korean AI companies. This includes the 5 companies designated as national AI champions by the Korean government last year—LG AI Research, SK Telecom, Naver Cloud, NC AI, and Upstage—as well as Naver, Kakao, Maum AI, and Krafton. For the HuggingFace analysis, academia (Seoul National University NLP) and an industrial AI user (Yanolja) were added as comparison benchmarks.

The Light and Shadow of Code Disclosure: Korean AI on GitHub

Based on GitHub API as of March 25, 2026.

There was a clear dividing line on GitHub. Five companies actively published code until March 2026, but Naver CLOVA and Kakao Brain have not uploaded new code since December 2024 and July 2024, respectively. However, the reasons the two companies stopped were completely different.

Naver integrated its CLOVA business into Naver Cloud in 2023, migrating its code to a new account. It merely changed its signboard while maintaining its open-source strategy. Kakao merged Kakao Brain into its headquarters' AI organization in August 2024, and after forming a partnership with OpenAI in the same year, decided not to create a new account to publish code. A company that used to create and publish its own technology chose instead to adopt external technology.

There were also differences among the active companies. Maum AI had the highest efficiency with 3,984 stars across 16 projects, but had no public models on HuggingFace. It is a strategy of publishing the code while keeping the models private. Game developer Krafton was similar. It actively updated 23 projects but did not enter HuggingFace.

Actual Model Usage: The Reason Behind the 254-Fold Difference

HuggingFace downloads showed a completely different picture from GitHub stars.

Based on HuggingFace API as of March 25, 2026.

First place went to LG AI Research, the AI research institute of LG Group. A single model, the 7.8 billion parameter Korean-English bilingual model EXAONE-3.5-7.8B-Instruct, recorded 330,000 downloads, while the total across 20 models reached 930,000. LG, which only first released a model in 2025, has secured the largest presence among the five government champion companies.

Telecommunications company SK Telecom ranked second with 720,000 downloads, but looking into the details reveals a different story. A single Korean text generation model released in 2021, ko-gpt2-base-v2, accounted for 94% of the total, with 680,000 downloads. The large model A.X-K1, released in January this year, garnered only 9,517 downloads. While the ranking is high, the number does not reflect its current competitiveness.

Downloads and community interest were also a different story. The AI startup Upstage had 100,000 downloads—one-ninth of LG's—but its likes were 2,481, nearly identical to LG's (2,607). The models developers pay attention to and the models they actually use were different. The game developer NCSoft, which received 5,209 stars on GitHub, ranked at the bottom on HuggingFace with 3,676 downloads and 291 likes. It was the company with the largest gap between code interest and actual model usage.

The Seoul National University NLP Lab, included as a comparison benchmark, ranked third with 550,000 downloads across 49 models. Academia, rather than a corporation, recorded more downloads than Naver. Yanolja recorded 38,000 downloads with just 2 models, meaning a travel platform rather than a specialized AI company surpassed Kakao (6,800 downloads) and NCSoft (3,676 downloads).

Code Disclosure and Model Adoption Are Different Games

When placing the numbers from the two platforms side by side, the rankings are reversed.

Stars: Cumulative. Downloads: Last 30 days.

NCSoft's GitHub stars were 4.4 times higher than LG's. However, in HuggingFace downloads, LG was 254 times higher than NCSoft. The same was true for Kakao. With 8,744 stars and 6,800 downloads, interest in the code did not translate into model adoption.

On the other hand, LG and Upstage had fewer GitHub stars with 1,171 and 1,426, respectively, but recorded 930,000 and 100,000 HuggingFace downloads. NCSoft uploaded 49 projects to GitHub but only had 9 models on HuggingFace, whereas LG had 22 GitHub projects and 20 HuggingFace models. The difference in what was published resulted in the difference in adoption.

Looking only at the 5 AI champions designated by the government, the gap was even more distinct. Based on HuggingFace downloads, LG (930,000) and SK Telecom (720,000) were overwhelmingly dominant, followed by Upstage (100,000). Naver Cloud (based on the Naver Cloud account) and NC AI had 289,000 and 3,676 downloads respectively, showing a maximum 254-fold difference even among the champions. Government designation did not necessarily mean technology adoption.

Neither government designation nor the number of GitHub stars revealed whether the published technology was actually being used. Only HuggingFace downloads held the answer.

This article was written based on the OpenIndex, created by News Epoch to quantitatively track the Korean AI open-source ecosystem.