4-Way Race Begins for Second Stage of Sovereign AI Foundation Model Project; Which Team Will Be Eliminated?

The second-stage evaluation for the "Sovereign AI Foundation Model" (Dokpamo) project, which selects South Korea's national artificial intelligence champions, is set to kick off.

Four teams—LG AI Research, SK Telecom, Upstage, and Motif Technologies—are competing, and only three of them will advance to the next stage.

Following the first-stage evaluation where four out of five teams survived, one team will now be eliminated.

This second evaluation differs in nature from the first, which focused on the technological completeness of the models and whether they were developed independently.

Capabilities as AI agents handling actual tasks, applicability to industrial sites, and openness and ecosystem expansion have emerged as key evaluation factors.

LG AI Research threw down the gauntlet with sheer "scale" and performance.

Its second-stage model, "K-EXAONE 2.0," boasts 750 billion (750B) parameters, making it the largest model in the country and more than three times larger than its first-stage model of 236B.

The average score across 24 benchmark evaluation metrics reached 70.1 points, over 10% higher than the 63.3 points of the first-stage model, while average performance in key coding and agentic coding metrics rose by 30%.

The company is also pursuing the open licensing of its model for commercial use alongside the expansion of industry-specific models.

SK Telecom placed its emphasis on industrial applicability and operational efficiency.

Its "A.X K2" model, featuring 688 billion (688B) parameters, applies a self-developed SGA (Sparse Gated Attention) architecture that selectively references only highly relevant information within long contexts.

It is expanding proof-of-concept cases across various industrial sectors, including steel and automotive parts manufacturing, national defense, and new drug development.

The company envisions enhancing global competitiveness by scaling its successor model up to the trillion-parameter level.

Upstage focused on lowering the barrier to entry for businesses adopting AI.

Its "Solar Open 2" adopts a Mixture of Experts (MoE) architecture that activates only 15 billion parameters during actual operation out of a total of 250 billion (250B).

It boasts cost-efficiency by being able to run on just two NVIDIA H200 GPUs while processing long contexts of up to 1 million tokens.

Motif Technologies, which started a month later than the other teams after being additionally selected last February, put technological independence at the forefront.

Its "Motif 3" model, with 314 billion (314B) parameters, was independently developed from architecture design to implementation.

The company explained that the model achieves comparable performance even with a total and active parameter scale smaller than equivalent Chinese open-source models.

As specialized overseas companies enhancing AI benchmark performance step up their offensives in the domestic market, the interpretation and fairness of benchmark scores have also emerged as key points of interest in the Dokpamo evaluation.

According to the AI industry, AI training specialists such as AfterQuery, Scale AI, and Mercoa engage in businesses that analyze tasks where models are vulnerable, provide post-training data, or process data tailored for Reinforcement Learning with Verifiable Rewards (RLVR).

RLVR is a method that improves performance by having AI repeatedly solve problems with clear, definite answers.

While it can rapidly improve a model's reasoning and coding performance, it can also spark controversies over excessive optimization tailored to specific benchmarks.

The fact that this second evaluation assesses not only simple benchmark scores but also agent task execution capabilities and industrial field applicability is interpreted as an effort to compensate for such limitations.

A procedure allowing ordinary citizens to directly use and evaluate domestic AI models is also being introduced.

The Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) selected 200 citizen evaluators by reflecting gender and age ratios based on resident registration demographics.

The evaluation panel will directly use the sovereign AI models of the four teams from the 8th to the 11th and score them through an absolute evaluation method.

The evaluation results will be reflected in the second-stage evaluation alongside benchmark and expert assessments.

The government plans to disclose detailed evaluation criteria and results within this month.

This evaluation goes beyond merely testing the technological level of Korean-style hyperscale AI.

Attention is focused on what will become the decisive criterion for advancing to the next stage: the performance of large-scale models, actual enterprise deployability, or user usability experienced by the public.
※ Please note: This article was translated by AI and may contain errors.