Understanding AI Foundation Model Development Projects in Korea

작성자:
A robotic hand reaching into a digital network on a blue background, symbolizing AI technology.

Korea’s main AI foundation model effort is a government-backed competition in which several teams build large language models from scratch and are cut down in rounds until two remain. The Ministry of Science and ICT (MSIT) supports the teams with funding, data and GPUs, and in return expects domestic technology and, in many cases, open-source release. This article explains how the program is structured, who has advanced, and how to read news about it.

What a foundation model is, and why Korea wants its own

A foundation model is a large model trained on broad data that can be adapted to many tasks, such as chat, search, coding help or image understanding. The Korea Times reported that the goal of the national project is for teams to develop large language models “using purely domestic technology to achieve global competitiveness and secure digital sovereignty.” Beyond text-only models, the teams are pursuing multimodal and omni models that handle text, images and audio and perform varied tasks.

How the project started

According to the Korea Times, five consortia were selected at the start: Naver Cloud, Upstage, SK Telecom, NC AI and LG AI Research. Each consortium includes partners. For example, the SK Telecom group includes Krafton, 42dot and Rebellions, and the LG group includes LG Uplus, LG CNS and FuriosaAI.

The same report lists the state support: combined government funding of 200 billion won, 10 billion won for dataset procurement from bodies such as the National Archives, Statistics Korea and the Korean Intellectual Property Office, 20 billion won worth of video training data, and GPU lending for three of the consortia (Upstage, NC AI and LG AI Research). Selected participants were to receive “K-AI Models” or “K-AI Companies” designation, and the ministry stressed that teams committed to open-source policies allowing commercial use by other companies.

How the elimination rounds work

SK Telecom’s own description of its role says the project undergoes evaluations every six months, with only two teams advancing to the end. It also states its own aims, including omni-modal models and performance at least 95% on par with global frontier models. These are the company’s targets, not guaranteed results.

The first round removed teams. The Korea Times reported that Naver Cloud and NC AI were removed in the first-stage evaluation, and that the ministry said Naver Cloud “failed to meet the criteria for developing a fully independent model architecture trained from scratch.” That report also says the first assessment focused on benchmark performance, expert review, user experience and technological originality. SK Telecom, LG AI Research and Upstage qualified, and a consortium led by Motif Technologies was added as a further participant.

Where the competition stands after the second stage

The Elec reports that in the second-stage results three teams advanced: LG AI Research with K-EXAONE 2.0, SK Telecom with A.X K2 and Upstage with Solar Open 2. Motif Technologies did not advance, even though its Motif 3 model was described as ranking tenth among top-performing models from AI companies worldwide.

The scoring used a 100-point scale: 40 points for benchmarks (25 from a global AI evaluation platform and 15 from the National Information Society Agency), 35 points for expert evaluation by ten experts, and 25 points for user evaluation involving 49 expert users and 185 members of the public. The report also says the government plans to eliminate one more team in the year-end final evaluation, leaving two, and that GPU support will increase from about 768 to 1,000 Nvidia B200 equivalents.

The Elec also highlights each finalist’s strengths, such as Upstage’s context window of up to 1 million tokens, SK Telecom’s math reasoning and LG’s 750-billion-parameter model. These are the outlet’s summaries of reported results, so check the ministry’s own announcements for exact figures.

What this means for businesses and developers

  • Open-source terms matter. The ministry said it wants open-source policies allowing commercial use by other companies, so read each model’s license before building on it.
  • Expect the field to narrow. Only two teams are meant to remain after the year-end evaluation, which may affect which models receive long-term state support.
  • Support is not a product. Funding and GPU access do not mean a model is ready for your use case. Test on your own tasks and languages.

How to read announcements about the project

  1. Note whether a claim is a company target, a ministry criterion or a measured result.
  2. Check which evaluation stage is being discussed.
  3. Look for the scoring weights, not only the headline rank.
  4. Separate a consortium’s brand from its model name, since a consortium contains several organizations.
  5. Treat benchmark numbers as dependent on the test used.

Common mistakes and what to verify

  • Treating selection as success. Early selection did not protect teams from elimination.
  • Using old lineups. The list of teams has changed more than once.
  • Quoting funding totals without context. The 200 billion won figure was reported at the start and may not describe later support.
  • Assuming every model is open. Check licenses individually.

Rules and standings change: confirm with the official body

This is general information, not investment, legal or procurement advice. Evaluation results, funding and licenses can change. For current facts, check announcements from the Ministry of Science and ICT and from the teams themselves.

Sources