A visitor uses SK Telecom’s A.X K1 model at the first presentation of the Independent AI Foundation Model Project, held on Dec. 30 at COEX in Gangnam, Seoul. (Image provided by SK Telecom)
SEOUL, Jan. 9 (Korea Bizwire) — South Korea’s race to build a homegrown artificial intelligence foundation model has drawn scrutiny, as SK Telecom moved this week to rebut claims that its flagship system borrows too heavily from a Chinese rival.
SK Telecom said on Thursday that its large language model, A.X K1, submitted to the government’s project to develop an independent national AI foundation model, is “structurally different” from models developed by China’s DeepSeek. The company was responding to industry speculation that A.X K1 shared key architectural features with DeepSeek’s latest systems.
In a statement, SK Telecom said A.X K1 uses a proprietary structure with 519 billion parameters, a configuration it described as unprecedented. The similarities cited by critics, the company said, relate only to so-called inference code—software used to run trained models efficiently—which is commonly adapted from publicly available sources and does not undermine a model’s originality.
SK Telecom emphasized that A.X K1 was trained “from scratch,” with all parameters randomly initialized and optimized by its own development team. In the AI industry, “from scratch” typically refers to building and training a model independently from the earliest stages, rather than fine-tuning an existing system.
This undated file photo provided by SK Telecom Co. shows the company’s headquarters in downtown Seoul. (Yonhap)
The debate reflects growing sensitivity around originality as South Korea pushes to establish sovereign AI capabilities. Lee Seung-hyun, a vice president at the AI company FortyTwoMaru, wrote on social media that while A.X K1 drew inspiration from certain architectural ideas used in other models, it delivered “value beyond from-scratch development” through large-scale optimization.
Similar questions have surfaced around other Korean firms participating in the government-backed initiative. Upstage has denied allegations that its model copied China’s Zhipu AI, saying it merely referenced inference code styles for open-source compatibility.
Naver Cloud has also faced criticism over claims that elements of its multimodal AI resemble Alibaba’s Qwen models, which the company said did not violate its commitment to independent development.
As competition intensifies to produce a credible domestic alternative to U.S. and Chinese foundation models, the disputes underscore how blurred the lines can be between inspiration, standardization and imitation in a field built on shared research—and how closely those distinctions are now being watched.
Kevin Lee (kevinlee@koreabizwire.com)
