Open-weight AI models made in South Korea

South Korea's government-backed sovereign AI foundation model project led four teams to release very large open-weight language models in 2026. This list covers those four, the two smaller Raon models that KRAFTON published in October and November, and the one speech model from a Korean developer in our catalog, and says which licenses carry conditions and which project is no longer maintained.

  • List updated November 11, 2026
  1. K-EXAONE 2.0 Largest

    LG's 750B mixture of experts with 37B active parameters, a 262K context and ten languages, under Apache 2.0. It is Korea's largest model and needs a large multi-GPU server.

    • LG AI Research
    • 750B (37B active)
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 518 GB
  2. A.X K2

    SK Telecom's text-only 688B mixture of experts with 33B active parameters, trained natively in FP8, under Apache 2.0. Its own table shows it strongest at math and Korean, and the card calls its agentic work moderate.

    • SK Telecom
    • 688B (33B active)
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 438 GB
  3. Motif 3

    Motif Technologies' 314B mixture of experts with 13.2B active parameters, built for agents and terminal work under the MIT license. It was the one team in the government project that did not advance to the next round.

    • Motif Technologies
    • 314B (13.2B active)
    • 256K tokens
    • MIT (Permissive)
    • from 206 GB
  4. Solar Open 2

    Upstage's 250B mixture of experts with 15B active parameters and a 1M-token context, for English, Korean and Japanese. It is Apache 2.0, but any model you build on it and share must have a name starting with Solar.

    • Upstage
    • 250B (15B active)
    • 1M tokens
    • Upstage Solar (Conditional)
    • from 157 GB
  5. Raon-1 mini preview

    KRAFTON's 28.8B mixture of experts with 4.3B active parameters, trained from scratch for Korean chat and agents under Apache 2.0, with a 262K context. A 17.6 GB NVFP4 build runs on one 32 GB GPU, and the card calls it a preview.

    • KRAFTON
    • 28.8B (4.3B active)
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 18 GB
  6. Raon-1 nano Newest

    KRAFTON's 3.1B on-device model (1.5B effective), distilled from the mini model for Korean and English with tool calling. The 4-bit builds are about 2 GB, and the license is Apache 2.0.

    • KRAFTON
    • 3.1B
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 1.98 GB
  7. Supertonic 3

    Supertone's 99M on-device speech model for 31 languages including Korean, which runs through ONNX Runtime. The OpenRAIL-M weights carry use restrictions, and the developer archived the project in September 2026.

    • Supertone
    • 99M
    • OpenRAIL-M (Conditional)
    • from 0.46 GB

How we chose

  • The developer is based in South Korea, according to the country on its developer page.
  • Open weights you can download today from the developer's official repository.
  • Scores come from the developers or from the ministry, so we order the language models by size instead of by score.
  • We read each license, because one model requires derivatives to be named Solar and the speech model has use restrictions.