A dispute over the qualifications of the elite teams participating in the government's "national AI model project" has spread from the Upstage and Naver elite teams to SK Telecom's elite team. The national AI model project is a core government initiative being pursued with the goal of building an AI model unique to Korea. SK Telecom pushed back against claims that its large-scale AI model "A.X K1," submitted to the first project briefing hosted by the Ministry of Science and ICT on the 30th of last month, is similar to the Chinese model "DeepSeek," saying it is "structurally different."
On the 8th, SK Telecom said, "A.X K1 is a model with an independent structure, with 519 billion identical parameters, a number not previously reported globally."
That day, some in the industry claimed that SK Telecom's A.X K1 resembles the detailed settings (parameters) of MLA (Multi-head Latent Attention) and MoE (Mixture of Experts), known as core architectures of the Chinese DeepSeek V3 model.
SK Telecom said, "A.X K1 is a model trained with all parameters randomly initialized in terms of weights, and is independently developed," adding, "The part said to be similar to DeepSeek is the 'inference code,' which is code provided for convenience when running an open model." It continued, "This is distinct from training code, which speaks to originality from scratch, and the AI industry does not regard it as an element that compromises from-scratch development." From scratch refers to building an AI model entirely from the very first stage.
Lee Seung-hyun, vice president of FortyTwoMaru, wrote on Facebook that day about A.X K1, "I previously posted that I wasn't sure whether SK Telecom's model was from scratch, and I apologize to the SK Telecom team who worked hard," adding, "SK Telecom touted the largest scale with 519 billion parameters, and it was trained 100% from scratch," and, "It borrowed DeepSeek's MLA and Gemma's dual norm, but optimized them through quantitative scaling and showed value beyond from scratch."
Earlier, Upstage also faced allegations that it copied the Chinese Zhipu AI's "GLM-4.5-Air" model. However, the company countered that it only referenced the inference code style and that it is a standard approach for open-source compatibility.
Naver Cloud likewise faced controversy that it borrowed the encoder and weights—core components of a Multimodal AI model—from China's Qwen 2.5 model by Alibaba, but explained that while it did strategically adopt an external encoder, this did not violate the from-scratch principle.
Some say the controversy grew because the government did not set standards for what constitutes from-scratch development in AI model building.