KRAFTON releases A.X-K2 Raon speech MoE model

KRAFTON, the South Korean company best known as the studio behind PUBG, has published A.X-K2 Raon Speech 21B-A3B, an open speech model that handles both synthesis and recognition. The release is available now on Hugging Face.

The model is built as a mixture-of-experts system with roughly 21 billion total parameters but only about 3 billion active per token. That design lets a large model stay comparatively cheap to run at inference time, since only a fraction of its weights fire on any given pass — an approach that has become common in recent large language and multimodal releases.

One model, two directions

Rather than shipping separate systems for each task, Raon Speech aims to cover both ends of the speech pipeline in a single any-to-any architecture:

• Text-to-speech (TTS) audio generation
• Automatic speech recognition (ASR), or speech-to-text

Combining these capabilities is notable because most open speech stacks still rely on distinct specialized models. A unified system can simplify deployment and, in principle, share learned representations between listening and speaking.

The repository lists an "other" custom license, so teams evaluating the model for production should read the terms closely before building on it. KRAFTON has not published detailed benchmark figures alongside the initial release, so real-world quality across languages and audio conditions remains to be tested by the community. Still, the entry of a major game company into open speech modeling is a sign of how broadly the field is expanding beyond the usual AI labs.