Krafton 开源发布自研 300 亿参数韩语大模型 Raon-1 mini 预览版（Apache 2.0 免费商用）

Krafton (259960.KS) is broadening the reach of its artificial intelligence technology by releasing a large language model it developed in-house free of charge.
The company said on the 21st that it had published a preview version of Raon-1 mini, a large language model with 30 billion parameters. Released ahead of the official launch, the model is distributed on Hugging Face, the global AI platform, under an Apache 2.0 license. Anyone can download it free of charge and use it not only for research but also for commercial services and product development.
Raon-1 mini is a 30-billion-parameter model that Krafton built on its own from design through initial training, and it supports Korean and English. To improve both performance and inference efficiency, Krafton combined a mixture-of-experts architecture with linear attention technology it developed itself. The approach activates only 4 billion of the 30 billion parameters needed for a given task, cutting the volume of computation. That speeds up responses and lowers inference costs, the company said.
The preview version of Raon-1 mini outperforms overseas models of comparable size and comes close to larger domestic models on some metrics, according to the company. It is particularly strong at natural conversation in Korean and responds quickly. It also performs well at following instructions precisely, such as specifying formats or adopting a persona. Instruction-following is essential for agent applications, in which a model handles multistep tasks on its own.
Krafton pointed to cost efficiency as the core strength of the model. As AI use spreads across industries, the company judged, performance relative to cost in real operating environments matters as much as performance alone. The preview model was therefore developed with a focus on Korean-language conversation, AI agent capability and fast response times.
Krafton plans to keep updating Raon-1 mini to improve its performance before releasing an official version. When that version is released, the company will also publish a technical report detailing the model's training process and final performance.
"The release of Raon-1 mini is the result of Krafton independently developing a competitive language model with AI research and development capabilities built up over more than five years," said Lee Kang-wook, Krafton's chief AI officer. "With a cost-effective large language model that raises performance while lowering the cost burden, we will widen the scope of AI use across a range of industries."
Krafton is also continuing research into the core technologies underpinning AI model development and has been presenting the results at leading international AI conferences. Its AI research organization has published 54 papers at major conferences including NeurIPS, ICML and ICLR, and nine main-track papers were accepted at NeurIPS 2026 this year.