KRAFTON has unveiled its proprietary AI foundation model, A.X. K2 Raon-Speech, on the global AI platform Hugging Face. This latest offering is the fruit of the second phase of research conducted as part of an elite task force led by SK Telecom, under the auspices of the Ministry of Science and ICT.

Boasting 21 billion parameters, this model has secured the top spot for overall Korean language performance and third place for English among open-source voice language models under 30 billion parameters. The evaluation spanned six distinct domains—including speech recognition, synthesis, comprehension, Q&A, and tool calling—utilizing 46 benchmarks such as LibriSpeech, K-SpokenSpeech, VoiceBench, K-VoiceBench, MMAU, KMMAU, API-Bank, and FunctionChat.

A.X. K2 Raon-Speech integrates KRAFTON's custom-trained speech encoder and codec into SK Telecom’s small language model to master the art of voice comprehension and generation. Traditional conversational systems have often suffered from the loss of emotion and intonation when converting a user's voice into text. However, KRAFTON's model preserves these nuances, resulting in significantly more natural responses.

Since launching the 'Raon' AI brand last April, KRAFTON has rolled out several tools, including Raon-Speech, the real-time conversational model Raon SpeechChat, the text-to-speech model Raon OpenTTS, and the vision encoder Raon VisionEncoder. KRAFTON CAIO Lee Kang-wook noted, "A.X. K2 Raon-Speech is the result of KRAFTON’s consistent research and advancement in AI foundation technology. We remain committed to refining these technologies to craft entirely new gaming experiences."

KRAFTON plans to leverage these newfound speech AI capabilities to evolve its Co-Playable Character (CPC) system. By enabling CPCs to better grasp user intent and respond with appropriate vocal cues, KRAFTON aims to create a more organic, immersive interaction between players and AI within their games.