The four models — Raon-Speech, Raon-SpeechChat, Raon-OpenTTS and Raon-VisionEncoder — span speech, voice and visual processing, reflecting Krafton's push into multimodal AI. The brand name draws from the native Korean word "라온," meaning joy, and is designed to capture what the company calls the essential pleasure of gaming through AI.
Raon-Speech, a 9-billion-parameter voice language model, claimed the top global ranking among open speech-language models under 10 billion parameters in both English and Korean, assessed across seven core tasks and 40 benchmarks.
Raon-SpeechChat, meanwhile, is the first real-time, full-duplex voice conversation model developed in South Korea, capable of interrupting and being interrupted mid-conversation.
Raon-OpenTTS, trained entirely on publicly available voice data, ranked among the world's best in blind listening tests against research-grade text-to-speech systems built on proprietary datasets.
Raon-VisionEncoder, the fourth model, outperformed Google's SigLIP2 in select visual recognition tasks and will be integrated into Krafton's broader independent AI foundation model project.
"The release of the Raon model series marks an important milestone in building our AI capabilities," Chief AI Officer Lee Kang-wook said.
"By open-sourcing large-scale training data and core models, we hope to contribute to the advancement of multimodal technology and the growth of South Korea's AI ecosystem."
Copyright ⓒ Aju Press All rights reserved.