Krafton on Oct. 21 unveiled a preview version of its self-developed 30B-class large language model (LLM), Raon-1 Mini.

Raon-1 Mini is an LLM Krafton developed directly, from model design to training from scratch. It supports Korean and English, and was released in preview form on Hugging Face ahead of its official launch.

It applies the Apache 2.0 license. Anyone can download it for free and use it not only for research but also for commercial services and product development.

The model adopts a mixture-of-experts (MoE) structure that activates only 4 billion of its total 30 billion parameters as needed. Krafton combined this with its own linear attention technology, focusing on improving response speed and inference cost efficiency.

Krafton said its internal evaluations showed Raon-1 Mini recorded higher performance than overseas models of a similar scale, and approached domestic large models on some metrics. It highlighted Korean-language dialogue, instruction following and response speed as key strengths.

Its instruction-following capabilities extend beyond format specification and persona settings to use as an AI agent handling multi-step tasks. Krafton said it used cost-to-performance in real usage environments as a key criterion, rather than simple benchmark scores, during development.

Raon-1 Mini also serves as the base LLM for Krafton's AI model brand Raon, launched in April. Earlier, Krafton released four models as open source: the speech language model Raon-Speech, the real-time voice conversation model Raon-SpeechChat, the text-to-speech conversion model Raon-OpenTTS and the vision encoder Raon-VisionEncoder.

Krafton plans to further improve Raon-1 Mini's performance and release an official version. At the official launch, it will also publish a technical report covering the training process and final performance.

Kangwook Lee (이강욱), Krafton's chief AI officer (CAIO), said, "Based on AI R&D capabilities accumulated for more than 5 years, we independently developed a competitive language model." He said, "Through an LLM that raises performance while lowering the cost burden, we will expand the scope of AI use across various industries."

Krafton's AI research organization has published 54 papers so far at major international AI conferences including NeurIPS, the International Conference on Machine Learning (ICML) and the International Conference on Learning Representations (ICLR). It said 9 main-track papers were accepted at NeurIPS 2026 this year.