Krafton has officially launched its proprietary AI model brand, 'Raon', and released four foundation models as open source on the global platform Hugging Face. The released models include a voice-enabled Large Language Model (LLM), a real-time voice conversation model, a Text-to-Speech (TTS) model, and a vision encoder. This release demonstrates Krafton's AI technical capabilities, covering the entire process from data collection to model training and performance evaluation.

The core model, 'Raon-Speech', is a 9-billion (9B) parameter voice-language model capable of both speech understanding and generation. It recorded the top global performance among open-source voice-language models under 10B parameters in both English and Korean. This result is based on a comprehensive evaluation of 7 core tasks and 40 benchmarks, including speech-to-text, text-to-speech, and voice-based Q&A.

Additionally, 'Raon-SpeechChat' is a real-time voice-language model featuring full-duplex communication, marking the first such release in Korea. The model demonstrated top-tier global performance across 13 tasks, including back-channeling, interruption handling, and response latency in full-duplex communication benchmarks.

'Raon-OpenTTS' is a text-to-speech model trained exclusively on public voice data. Krafton collected and refined the data internally and has provided the entire training dataset. In blind human evaluations, the model proved to have top-tier performance compared to private, research-only global TTS models.

'Raon-VisionEncoder' is a vision encoder that converts images into information understandable by AI. This model was trained from scratch using only public data without relying on pre-trained models. In some visual recognition tasks, it outperformed Google's vision encoder model (SigLIP2), and in other tasks, it maintained over 90% of SigLIP2's performance, proving its competitiveness. This technology will be utilized in future 'proprietary AI foundation model' projects.

Lee Kang-wook, Krafton's Chief AI Officer (CAIO), stated, "The release of the Raon model series is an important milestone in the process of accumulating AI technical capabilities." He added, "By sharing large-scale training data and core models as open source, we hope researchers and developers can utilize them freely and contribute to the advancement of multimodal technology and the growth of the domestic AI ecosystem." Krafton plans to continuously strengthen its global AI technical competitiveness centered on Raon.

Krafton introduced its personal AI assistant 'KIRA' last year, and last month, it strengthened its technical competitiveness by releasing 'Terminus-KIRA', a technology that improves AI agent performance, as open source. Furthermore, the company is expanding new gaming experiences through AI-based interactive content, such as co-playable characters, and plans to continue leading technological innovation by advancing its AI models and agent technologies.