Krafton has launched its artificial intelligence (AI) model brand Raon on April 1, unveiling four AI models as open-source.
Raon, derived from a pure Korean word meaning joy, embodies Krafton’s philosophy of harnessing AI technology to create fundamental enjoyment in gaming.
On this day, Krafton released four open-source models on Hugging Face: a voice-enabled large language model (LLM), a real-time voice conversation model, a text-to-speech (TTS) model, and a vision encoder.
The voice-enabled LLM, Raon-Speech, extends text-centric language models to comprehend and generate speech. With 9 billion parameters, it has achieved top performance globally among voice language models under 10 billion parameters for both English and Korean.
Raon-SpeechChat, the real-time voice conversation model, is a pioneering interactive voice model in Korea.
It features real-time bidirectional communication technology, allowing seamless interruptions between users and the model during conversations.
The TTS model, Raon-OpenTTS, is trained exclusively on publicly available voice data.
Krafton has collected and refined previously challenging-to-use data for public release. They’ve also made the entire training dataset accessible, enabling anyone to replicate the training environment.
The Raon-VisionEncoder converts images into AI-comprehensible information.
In certain visual recognition tasks, it has even outperformed Google’s vision encoder model SigLIP2. This technology is set to be utilized in the government’s Independent AI Foundation Model project.