Korea cuts one of four sovereign AI teams, and usability did the cutting

South Korea's Ministry of Science and ICT announced on the 18th, at a briefing at the Government Complex Seoul, that three of four teams cleared the second stage of the sovereign AI foundation model project: Upstage, SK Telecom and LG AI Research ^[2][3]. Motif Technologies finished last and was eliminated after low marks on AI service usability, even though its Motif 3 model scored highest among the teams on the Artificial Analysis Intelligence Index ^[3][11][15].

The scoring is the story. The three assessment areas carry 25 points for user assessment, 40 for benchmarks and 35 for expert review, a total of 100 ^[6][7][8][9]. The user assessment produced the widest spread: a 5-point gap between top and bottom on 25 available points, or 20 percent of the bucket, against 4 points on 40 in benchmarks and 2.4 points on 35 in expert review, under 7 percent ^{[6][7][8][10]}. In other words, the smallest scoring bucket did the most separating, and the benchmark round was where the teams sat furthest below the available marks, averaging 22.5 of 40 ^[7][c10b].

The usability test was run with 49 professional users, including AI startup executives, and 185 members of the general public, 234 people in total ^[12][13]. Second Vice Minister Ryu Je-myung said at the briefing that "the top performer differed by area, and no single company ranked first consistently," and that Motif "clearly delivered outstanding results, with a very high score on the Artificial Analysis Intelligence Index (AAII), but it was rated lower than the other companies on usability and applicability" ^[14]. Note that the AAII was itself worth 25 points in the second stage, which means it sat inside one of the three buckets rather than alongside them, since those already total 100 ^[15][16].

The survivors all have benchmark stories too. Upstage's Solar Open 2 secured a context window of up to 1 million tokens, which the ministry's account puts on par with Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 Sol for long-text processing ^[17]. SK Telecom's A.X K2 reached a gold-medal-level score on this year's International Mathematical Olympiad problems ^[18]. LG AI Research's K-EXAONE 2.0, at 750 billion parameters, the largest in the country, ranked ninth in the world on minimizing hallucination ^[19].

What separated them was deployment. Upstage's model was adopted for "Yeon.Ye.In," a system that reviews the national R&D budget by analyzing projects and budget data for overlaps and duplication and drafting documents ^[21]. An EXAONE-based materials discovery system was applied to cosmetic ingredient development at LG Household & Health Care ^[22]. SK Telecom is putting A.X K2 into manufacturing AI agents with KG Steel and the auto parts maker Conec ^[23]. The competition is widening into public administration, manufacturing, defense and education ^[20].

Two of the three will be selected as early as the end of this year, roughly four months out ^[5][29]. Watch what each team does with that window, because the signals are already public: Upstage plans to add Asian languages ^[25], SK Telecom says it will produce outcomes the whole Korean AI ecosystem can use ^[26], and LG AI Research is aiming K-EXAONE at frontier-class expert work, with co-head Yim Woo-hyung saying "it is difficult to close the technology gap with global Big Tech using small models alone" ^[24][27]. Enterprise buyers should read the reference deployments as the real shortlist, since that is the axis the state just scored on.