Look, let's be honest—KRAFTON has just dropped a preview version of its homegrown 30 billion-parameter large language model, Raon-1 mini, on August 21st. The model is now available as open-source via Hugging Face under the Apache 2.0 license, meaning anyone can grab it for free. KRAFTON notes that Raon-1 mini isn't just for academic pondering; it's ready to be put to work in commercial services and product development.

The Raon-1 mini preview is a 30 billion-parameter LLM where KRAFTON handled everything from the initial design down to training from scratch. It speaks both English and Korean. To balance high performance with inference efficiency, the developer crafted a proprietary model architecture. This setup combines a Mixture of Experts (MoE) structure—activating only 4 billion parameters out of the total 30 billion when needed—with a custom linear attention technology to keep response times snappy and inference costs low.

The preview model holds its own against similarly sized overseas offerings and approaches the level of major domestic models in several metrics. It boasts solid Korean conversational skills and response speeds, alongside the prompt-following chops needed for formatting and persona setting—making it a handy tool for multi-step agent tasks. KRAFTON realised that as AI adoption ramps up, cost-performance efficiency is everything, which is why the team aimed for an economical, budget-friendly model.

KRAFTON plans to continuously update Raon-1 mini to iron out performance wrinkles before launching a full, official version down the line. When that official rollout happens, a detailed technical report outlining the training journey and final benchmarks will accompany it. Lee Kang-wook, KRAFTON's Chief AI Officer (CAIO), stated, "Releasing Raon-1 mini is a testament to KRAFTON's independent capability to build a competitive language model, backed by over five years of accumulated AI research and development. We aim to expand AI adoption across various industries with a cost-effective large language model that ramps up performance while dialing down the financial burden."

Back in April, KRAFTON launched the AI model brand 'Raon', already open-sourcing four distinct AI models including Raon Speech, Raon Speech Chat, Raon Open TTS, and Raon Vision Encoder. This newly unveiled Raon-1 mini preview will serve as the foundational bedrock for KRAFTON's future AI models. Through this development, the company has secured the technological know-how to tailor performance and inference efficiency to whatever an AI use case demands. Furthermore, fueling this development with foundational research, the team has published 54 papers in top-tier international AI conferences such as NeurIPS, ICML, and ICLR, notably securing nine main-track paper acceptances at NeurIPS 2026 this year.