Companies participating in Korea’s independent AI foundation model project, known as Dokpamo, are making an all-out push to secure agentic AI capabilities ahead of a second-stage evaluation in August.

Dokpamo developers are focusing on strengthening agentic AI performance as global large language model development shifts. As LLM use expands beyond simple Q&A into actual work execution, orchestration capabilities that call tools such as external search, databases and code execution have emerged as a key competency. Global leaders such as OpenAI, Anthropic and Google have highlighted agent and tool-calling benchmark figures when announcing new models since late 2024.

Domestic companies participating in Dokpamo also appear to be moving in line with this trend.

LG AI Research is focusing on advancing orchestration capabilities in developing the second-round evaluation version of its K-EXAONE. While the first round centered on basic language skills, the second round has shifted the development focus to agent task execution capabilities. LG AI Research said this direction reflects potential use in affiliate and client AX projects. It said once companies start using an orchestrator and see productivity gains, it is not easy to move away.

It is also working in parallel on tuning to deploy EXAONE in affiliates’ and partners’ day-to-day operations. The institute said it tried a similar effort 3 years ago but failed after 1 year as each stage of model development took months, including receiving data, reviewing training data quality, and evaluation and training. Based on that experience, it said it has now raised both development speed and quality.

An LG AI Research official said, "Some agent tasks are also possible," adding, "In the final stages of Dokpamo development, there were disappointing areas in terms of speed, so several labs were put into reinforcement work and substantially raised orchestration capabilities over about 2 weeks."

Upstage recently unveiled its competing Dokpamo model, the Solar Open2 Preview, and stressed that it made progress in agent capabilities. It also disclosed agent capability figures. Upstage said the Solar Open2 Preview achieved 98 percent agent performance on the Tau2 benchmark. Upstage said this surpasses DeepSeek V4 Pro at 96.2 percent and is comparable to Anthropic Fable5 at 98.5 percent. It said the figures were independently measured by AI performance evaluation group Artificial Analysis.

Upstage CEO Sung-hoon Kim (김성훈) said, "The first model could be used for media proofreading, Q&A and search, but it could not call tools for agent use," adding, "The model to be released in June can be sufficiently used for agent use." He added, "AI has become a national strategic asset, not a service or tool," and stressed, "If they want, any country can cut it off, so we must raise our own technology as much as we can."

SK Telecom is also developing a new model for the second-round evaluation, A.X K2, with a focus on agent capabilities. It has focused development on coding and agent capabilities since the previous version, A.X K1. The company said K2 will have agent performance that can be used across multiple domains and scenarios. It is also reflecting agent research results from a research team led by Seoul National University professor Seung-won Hwang (황승원).

An SKT official said, "We are newly building a model to submit for the second-stage evaluation, and it will have agent performance that can be used in various forms."

Motif Technologies has also set its goal for a 300B-class LLM under development for the second round as advancing agent performance. Motif stressed as a differentiator that it uses a model it designed independently rather than using foreign open-source architectures. It said it aims to prove competitiveness in agent performance as well.

Motif Technologies CEO Jeong-hwan Lim (임정환) said, "As the development goals of global LLMs such as Claude are shifting toward the ability to call tools that agents will use, Dokpamo is also aligning with that," adding, "Motif is also aligning its development goals accordingly."

Separately, the Ministry of Science and ICT will conduct the second-stage Dokpamo evaluation for the 4 companies simultaneously in early August. The submission deadline is by the end of June for LG AI Research, SKT and Upstage, and by the end of July for Motif Technologies.