Second Phase of the DOKPAMO Evaluation Begins... Elimination Looms for "National AI" Teams

LG AI Research Institute, SK Telecom, Upstage, and Motif in a Four-Way Competition

Citizen Evaluators Begin Their Assessments

The second phase evaluation of the so-called "National AI" project, officially known as the Independent AI Foundation Model (DOKPAMO) project, is about to begin. Four teams—LG AI Research Institute, SK Telecom, Upstage, and Motif Technologies, which joined through an additional call for applications—will compete, with only three advancing to the next phase.

According to the IT industry on August 9, a panel of citizen evaluators will directly use the independent AI models developed by these four teams and rate them using an absolute evaluation method from August 8 to August 11. Their assessments, along with benchmark and expert evaluations, will be reflected in the results of the second phase evaluation.

Deputy Prime Minister and Minister of Science and ICT Background Hoon, Vice Chairman of the National AI Strategy Committee Lim Moon-young, and Hajungwoo, Chief Secretary for AI Future Planning, are posing for a commemorative photo at the first presentation of the "Proprietary AI Foundation Model" project held on the afternoon of the 30th at COEX in Gangnam-gu, Seoul. Photo by Yonhap News

Previously, the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) selected 200 citizen evaluators, taking into account gender and age ratios based on resident registration statistics.

The reason for incorporating the citizen panel’s assessments in the second phase is the increasing importance of evaluating the models’ competencies for real-world use. This round of evaluation focuses on three main factors: the AI agents’ ability to handle actual tasks, applicability in industrial settings, and openness and potential for ecosystem expansion.

Companies participating in the project are sequentially releasing their models for the second evaluation round.

For example, LG AI Research Institute’s phase-two model, "K-Exaone 2.0," has been substantially scaled up. With a parameter count of 750 billion, it is more than three times larger than the previous 236 billion-parameter model, making it the largest domestically. The average score across 24 benchmark indicators is 70.1 points, which is over 10 percent higher than the first model’s 63.3 points. The average performance on key coding and agentic coding metrics improved by 30 percent. The institute is also expanding model licensing for commercial use and developing industry-specialized models.

SK Telecom has unveiled a 688 billion-parameter model, "A.X K2." This model incorporates an in-house developed Sparse Gated Attention (SGA) structure, which selects and references only highly relevant information from long-context data. It is being validated across multiple industries, including steel and auto parts manufacturing, defense, and new drug development. SK Telecom plans to expand subsequent models to a parameter scale reaching trillions.

Upstage has focused on boosting model efficiency to lower the barriers for enterprise AI adoption. Its "Solar Open 2" model uses a Mixture-of-Experts (MoE) structure, with only 15 billion out of the total 250 billion parameters activated during operation. This allows it to process up to 1 million tokens of lengthy data while being able to run on just two NVIDIA H200 GPUs, delivering high cost efficiency.

Motif Technologies, newly selected in February, highlights its technological independence. Its "Motif 3" model, with 314 billion parameters, was developed entirely in-house—from architecture design to implementation. According to Motif, the model achieves comparable performance to Chinese open-source models in the same class, despite having fewer total and active parameters.

This second-phase evaluation will assess not only benchmark scores, but also the models’ ability to perform agent tasks and their applicability in industry. This approach is intended to offset the limitations of benchmark-based performance optimization, which can sometimes inflate benchmark scores compared to real-world capacity.

The government plans to release the detailed evaluation criteria and results later this month.

This content was produced with the assistance of AI translation services.