KRAFTON AI, the artificial intelligence division under South Korean game company KRAFTON, released a preview version of the Raon-1 mini this week. This is a model trained from scratch, and the official version as well as a larger follow-up model are in development. In the competitive field of base models dominated by giants, a game company bringing its own model shows a bold ambition to cross over into other domains.
30B-A4B's "small but full"
Raon-1 mini adopts a mixed expert (MoE) architecture and introduces linear attention, supporting speculative decoding. The overall parameter count is marked as 30B-A4B—that is, approximately 30 billion parameters, with about 4 billion activated at a time. It focuses on Korean while also covering English, and is open-sourced under Apache 2.0 license. The official website also offers a limited-time free API trial.
The report given by KRAFTON AI is straightforward: the preview version is on par with Meta’s Muse Glimmer 30B in terms of programming and agent tasks, but it runs faster, uses less memory, and can engage in natural conversations. In other words, it aims to create a gap in efficiency and cost-effectiveness within the same scale.
Why do game companies focus on models?
It’s not hard to understand why a game studio that started its career with *Genshin Impact* is personally working on modeling. Game development itself has strong demands for story generation, NPC dialogue, art, and code support. Developing games in-house can reduce costs and give control over the process; releasing models as open source for testing is also a cheap way to build a good reputation and attract developers and an ecosystem.
As giants like Meta and Google build larger and larger base models, KRAFTON chose to start from a 30B level that is “sufficient and easy to deploy,” likely targeting users who want Korean localization capabilities but do not want to be tied to these giants. Whether Raon-1 mini can live up to its promises will be determined by the performance of the official version and the larger models, and it also marks the first step toward the idea that “game studios can also build a solid foundation.”