Krafton said on Wednesday it launched its artificial intelligence (AI) model brand Raon and released as open source on the global platform Hugging Face a large language model (LLM), a real-time speech conversation model, a text-to-speech (TTS) model and a vision encoder.

Raon was inspired by a native Korean word meaning joy. Krafton carried out in-house the entire foundation-model development process, including data collection, model training and performance evaluation.

The released models are Raon-Speech, Raon-SpeechChat, Raon-OpenTTS and Raon-VisionEncoder.

Raon-Speech is a speech language model capable of understanding and generating speech by extending a text-based language model. With 9 billion parameters, it achieved the top performance in English and Korean among publicly available speech language models with 10 billion parameters or fewer. Krafton said the result was based on a combined evaluation across 7 tasks and 40 benchmarks, including speech-to-text, text-to-speech and speech-based question answering.

Raon-SpeechChat applies full-duplex real-time two-way communication technology that allows users to interrupt during conversation. Krafton said it recorded top-tier average performance across 13 tasks, including backchannel responses, interruption handling and response latency, in 3 benchmarks for evaluating two-way communication models.

Raon-OpenTTS is a text-to-speech model trained on public speech data. Krafton said it collected and refined 일부 data and released it, and will also provide the full training dataset as open source.

Raon-VisionEncoder is a vision encoder that converts images into information that AI can understand. Krafton said it trained the model itself without using a pre-trained model. It said the model outperformed or showed more than 90 percent performance compared with Google vision encoder model SigLIP2 in some visual recognition tasks. The technology will be used in the "proprietary AI foundation model" project.

Kangwook Lee (이강욱), Krafton's chief AI officer (CAIO), said, "The release of the Raon model series is part of the process of accumulating AI technological capabilities." He added, "By sharing training data and core models as open source, we hope researchers and developers can use them and that it will contribute to the growth of the domestic AI ecosystem."

Krafton last year introduced its personal AI assistant KIRA. Last month it also released as open source the Terminus-KIRA technology to improve AI agent performance.