Upstage, LG and SKT advance to final stage of proprietary AI model competition

When the results were announced at the Government Complex in Jongno-gu, Seoul, the number of national flagship AI candidates fell from four teams to three. Upstage, LG AI Research and SK Telecom passed the second-stage evaluation of the proprietary AI foundation model project, while Motif Technologies failed to advance to the next stage.

This evaluation was not decided by a single performance record. It combined domestic and international benchmarks, assessments by external experts, and a user evaluation involving AI startup executives and members of the public. The Ministry of Science and ICT said the differences between the four teams were narrow enough for the rankings to shift depending on the evaluation, and it did not disclose the top-ranked team for each detailed metric.

Each team also targeted different use cases. Upstage's ‘Solar Open2’ was developed as a 250B model, combining some inherited weights from its predecessor with random initialization. It was designed for agent applications such as office productivity, document work and coding, as well as long-context reasoning. LG AI Research's ‘K-Exaone 2.0’ expanded the existing model through an upcycling approach, increasing its scale by more than threefold while improving long-document retrieval and safety. SK Telecom's ‘A.X K2’ has 688B parameters and uses a mixture-of-experts architecture that reduces the number of parameters activated during inference to around 33B.

Motif Technologies, which was eliminated, completed development of ‘Motif3’, a 314 billion-parameter model built from scratch, along with modules for three modalities. It scored 47 points on the Intelligence Index from global AI evaluator Artificial Analysis, the highest among the four teams, but received relatively low scores in the assessments of real-world usability and applicability, resulting in its elimination. The outcome shows that high model performance alone does not guarantee advancement to the final stage.

So what changes in practical terms? The three teams will each receive 1,000 B200 GPUs, high-performance graphics processing units, allowing them to further develop and enhance their models through the end of the year. Two of them will be selected as the country's flagship AI models early next year. However, the criteria and weighting for the third-stage evaluation will be finalized after an objection process and consultations among the three teams, making it difficult to predict the final rankings at this point.

The government has defined the project not merely as a contest to determine a winner between two teams, but as an initiative to strengthen the competitiveness of South Korea's AI ecosystem. Against the backdrop of moves by leading countries such as the United States to turn AI models into strategic assets, it also plans to announce a 2027 project plan for top-tier proprietary AI models after budget negotiations.