[Edaily Reporter Kim Hyun-ah ] KRAFTON(259960)unveiled a preview version of “Raon-1 mini,” a massive language model (LLM) with 30 billion (30B) parameters developed in-house, on the 21st.

The company handled everything from model design to training in-house, focusing on Korean-language dialogue, the execution of instructions by Agent AI, and inference efficiency.

The Raon-1 mini preview has been released on the AI model sharing platform Hugging Face under the Apache 2.0 license. Anyone can download it for free and use it for research and the development of commercial services.

This model employs a Mixed Expert (MoE) architecture, which activates only 4 billion of its total 30 billion parameters during the inference process. It also incorporates KRAFTON’s proprietary linear attention technology. The design aims to reduce computational load and improve response speed and inference efficiency by activating only the necessary parameters.

KRAFTON stated that Raon-1 Mini demonstrated competitive performance compared to overseas models of similar scale and, in some evaluation metrics, came close to large domestic models. The company also identified the ability to follow user instructions—such as Korean dialogue, formatting, and role-setting—as a key development goal. This ability to carry out instructions is utilized in the implementation of Agent AI agents capable of handling multi-step tasks.

KRAFTON explained that it considered not only model performance but also the inference costs required for actual service operation. This is because, when operating an AI service, the computational resources and costs required to process requests are just as important as the model’s performance.

This version is a preview model released ahead of the official launch. KRAFTON plans to release the official version after improving its performance and will also publish a technical report detailing the training process and evaluation results.

Lee Kang-wook, KRAFTON’s Chief AI Officer (CAIO), stated, “The release of Raon-1 Mini is the result of KRAFTON’s independent development of a language model, built on the AI R&D capabilities we have accumulated over more than five years,” adding, “We will expand the use of AI across various industries.”

Last April, KRAFTON launched its AI model brand “Raon” and released four types of AI models as open source: a speech language model, a real-time speech dialogue model, a speech synthesis model, and a vision encoder. The research team has published 54 papers to date at international AI conferences such as NeurIPS, ICML, and ICLR, and nine of their papers were accepted into the main track at this year’s NeurIPS 2026.