Krafton (259960.KS), the South Korean gaming company behind PUBG, announced Wednesday that it has launched an artificial intelligence model brand called "Raon" and released four AI models as open source on Hugging Face: a speech-enabled large language model (LLM), a real-time voice conversation model, a text-to-speech (TTS) model, and a vision encoder.
The name Raon derives from a pure Korean word meaning "joy," with its English spelling drawn from letters in "KRAFTON." The brand reflects Krafton's philosophy of using AI technology to create the fundamental joy of gaming.
Through this release, Krafton demonstrated its technical capability to independently handle the entire foundation model development process, from data collection to model training and performance evaluation. The company plans to further strengthen its global AI competitiveness centered on the Raon brand.
The four models released are Raon-Speech, Raon-SpeechChat, Raon-OpenTTS, and Raon-Vision Encoder.
Raon-Speech is a speech language model that extends text-based language models to enable speech understanding and generation. With 9 billion (9B) parameters, it recorded the top global performance in both English and Korean among open speech language models with fewer than 10 billion parameters. The ranking was based on a comprehensive evaluation of seven core tasks and 40 benchmarks — including automatic speech recognition, text-to-speech, and speech-based question answering — with average rankings across tasks weighted equally.
Raon-SpeechChat is a speech language model that applies real-time full-duplex communication technology, allowing both the user and the model to freely interrupt each other during conversation. It is the first real-time full-duplex voice model announced in South Korea. Across three full-duplex model evaluation benchmarks covering 13 key tasks — including backchanneling, interruption handling, and response latency — it recorded top-tier global performance based on average task rankings.
Raon-OpenTTS is a text-to-speech model trained exclusively on publicly available speech data. Some datasets that were previously difficult to use were directly collected and refined by Krafton and released publicly, along with the full training data, enabling anyone to reproduce the training under identical conditions. In blind human evaluations comparing the naturalness of two speech outputs, it demonstrated top-level performance compared to global research TTS models trained on proprietary data.
Raon-Vision Encoder is a vision encoder that converts images into information that AI can understand. When combined with a language model, it can process visual information. The model was trained from scratch using only publicly available data, without relying on pretrained models. It surpassed Google's flagship vision encoder model, SigLIP2, on certain visual recognition tasks. On other tasks, it demonstrated over 90% of SigLIP2's performance, proving its competitiveness. The technology will also be used in Krafton's proprietary AI foundation model project.
"The release of the Raon model series is an important milestone in the process of building our AI technology capabilities," said Lee Kang-wook, Krafton's Chief AI Officer (CAIO). "We hope that by sharing large-scale training data and core models as open source for researchers and developers to freely use, we can contribute to the advancement of multimodal technology and the growth of Korea's AI ecosystem."