Naver, NC AI knocked out of Korea‘s national AI model project

(Ministry of Science and ICT)

Naver Cloud and NC AI have failed to advance past the first round of South Korea’s flagship sovereign AI foundation model project, after evaluators ruled that Naver’s model did not meet requirements for technological independence, widening the number of eliminated teams from one to two.

On Thursday, Ministry of Science and ICT said that LG AI Research, Upstage and SK telecom advanced to the second stage of the project, which aims to develop a nationally representative AI foundation model.

The first-round evaluation consisted of benchmark testing worth 40 points, expert reviews accounting for 35 points and user assessments carrying 25 points. The ministry said it conducted a comprehensive review covering model performance, real-world applicability, cost efficiency and potential spillover effects on the domestic and global AI ecosystem.

LG AI Research topped all three evaluation categories with a total score of 90.2 points, well above the five-team average of 79.7 points. Scores for the other shortlisted teams were not disclosed. NC AI was eliminated after its combined score across benchmark, expert and user evaluations fell short of the cutoff.

The ministry said it plans to launch an additional call for applications within the first half of the year to select one more elite team, ultimately forming a four-team competitive structure. “The goal of the project is to give all participating companies an opportunity to leap to a global level,” the ministry said, adding that administrative procedures would be expedited to move forward quickly.

Naver’s elimination, in particular, sent shockwaves through the IT industry, as the company had been widely viewed as a strong contender. Naver is regarded as a symbol of Korea’s sovereign AI push and is the developer of the country’s first generative AI model series, HyperCLOVA. However, controversy arose after it was revealed that the model used an encoder and weights from Alibaba’s open-source Qwen model.

Naver argued that the Alibaba encoder was a replaceable module adopted for global compatibility and accounted for only a small portion of the overall system. While such strategic use of open-source components is common in the global AI industry, the ministry said the model failed to meet independence standards given its intended use as national infrastructure, including potential applications in defense, diplomacy and security.

Naver said it respects the ministry’s decision and effectively ruled out participating in the additional call. Kakao, which had also been mentioned as a possible reapplicant, said it would not seek re-entry either, a move industry watchers attribute to concerns over brand image.

The IT sector expects both Naver and Kakao to step back from government-led projects for the time being and refocus on strengthening their core technologies and developing new businesses.