KRAFTON Announces AI Model Brand "Raon"

- Four models released: voice-enabled LLM, real-time voice interaction, TTS, and vision encoder. - Vision encoder trained from scratch without using existing models. KRAFTON Inc. (CEO Chang-han Kim) has announced a new AI model brand, "Raon." Four models – a voice-enabled large language model (LLM), a real-time voice interaction model, a text-to-speech (TTS) model, and a vision encoder (image feature extraction model) – have been released as open-source on the global platform "Hugging Face." "Raon" is a brand name inspired by a Korean word meaning "joy" and also incorporates some letters from KRAFTON's company name. It reflects the company's philosophy of creating the essential joy of games through AI technology. KRAFTON has released four models, "Raon-Speech," "Raon-SpeechChat," "Raon-OpenTTS," and "Raon-VisionEncoder," backed by its technical capabilities to execute all processes of foundation model development, from data collection to model training and performance evaluation. Moving forward, the company plans to further strengthen its global AI technology competitiveness with "Raon" at its core. * "Raon-Speech" is a voice language model that extends text-centric language models to enable understanding and generation of speech, securing a scale of 9 billion parameters. Among publicly available voice language models under 10B, it recorded the highest global performance in both English and Korean. This result is based on a comprehensive evaluation of 7 key tasks, including speech recognition, speech synthesis, and voice-based question answering, across 40 benchmarks, reflecting the average rank of each task with equal weighting. * "Raon-SpeechChat" is a voice language model that employs real-time full-duplex communication technology, allowing users and the model to interact with natural interruptions during conversations. It is the first real-time full-duplex voice model announced in Korea and has shown top-tier global performance based on the average ranking of 13 key tasks, including interjections, interruption handling, and response delay time, across three evaluation benchmarks for full-duplex communication models. * "Raon-OpenTTS" is a text-to-speech model trained using only publicly available voice data. KRAFTON itself collected and refined some data that was previously difficult to utilize and has released it along with the entire training dataset, making it possible for anyone to reproduce the training in the same environment. In blind evaluations by humans, it achieved top-tier performance even when compared to global research-oriented TTS models using non-public data. * "Raon-VisionEncoder" is a vision encoder that converts images into AI-understandable information. When combined with a language model, it enables the processing of visual information. Its distinctive feature is that it was trained from scratch using only public data, without employing pre-trained models. In some visual recognition tasks, it recorded results surpassing Google's representative vision encoder model "SigLIP2," and in other tasks, it showed over 90% performance compared to SigLIP2. This technology is also planned to be utilized in KRAFTON's "Proprietary AI Foundation Model" project. ■Comment from KRAFTON's CAIO (Chief AI Officer) "The release of the 'Raon' model series is an important milestone in our steady accumulation of AI technology. By sharing large-scale training data and core models as open source, we aim to create an environment where researchers and developers can freely utilize them, and we expect to contribute to the development of multimodal technology and the growth of Korea's AI ecosystem." KRAFTON possesses comprehensive development capabilities across all AI technologies, from foundation model design to AI agents and the application of CPC (Co-Playable Character) technology to games to create new gaming experiences. Based on this, KRAFTON unveiled the personal AI assistant "KIRA" in 2025 and open-sourced the "Terminus-KIRA" technology to enhance AI agent performance in March 2026. Furthermore, it is working to expand new gaming experiences through AI-based interactive content like CPC. Moving forward, the company plans to continue advancing AI models and AI agent technologies, leading technological innovation.