The current state of voice interaction with artificial intelligence often feels like a relay race. A user speaks, a Speech-to-Text engine transcribes the audio into text, a Large Language Model processes that text to generate a response, and finally, a Text-to-Speech engine converts that response back into a synthetic voice. This cascaded pipeline is the industry standard, but it is fundamentally flawed, introducing latency at every handoff and stripping away the emotional nuance of the human voice. The developer community has been waiting for a native, end-to-end multimodal approach that treats audio not as a transcription task, but as a primary language.

The Architecture of A.X K2 Raon-Speech

Krafton has entered this arena with the release of A.X K2 Raon-Speech, a bilingual Korean-English voice AI model designed to bypass the traditional pipeline. The technical foundation of the model is a sophisticated hybrid of existing high-performance components. At its core, it utilizes the A.X K2 Light 20B-A3B text backbone developed by SKT. To handle the auditory dimension, Krafton integrated a proprietary AuT encoder based on Qwen3-ASR and a Mimi-family neural speech codec. This combination allows the model to perceive and generate audio natively rather than relying on intermediate text representations.

From a structural standpoint, the model employs a Mixture of Experts (MoE) architecture. While the total parameter count stands at 21.2B, the MoE design ensures that only approximately 3.5B parameters are activated during any single inference pass. This optimization allows the model to maintain the knowledge capacity of a larger network while operating with the computational efficiency of a much smaller one. The model is provided under the CC BY-NC 4.0 license, making it available for research and non-commercial prototyping.

Deployment requirements are significant due to the model's precision. The BF16 weights occupy approximately 42.4GB of memory, which means a single 80GB GPU or a multi-GPU configuration is recommended for stable operation. Beyond the hardware, Krafton has implemented a strict ethical guideline: users must obtain explicit consent when using or generating the voices of identifiable individuals.

Beyond Transcription to Native Multimodality

What distinguishes A.X K2 Raon-Speech from previous open-weights voice models is its refusal to be a single-purpose tool. Most models in this size class are either excellent at transcription or proficient at synthesis. Raon-Speech collapses these functions into a single session. It supports Speech-to-Text (STT), Text-to-Speech (TTS), SpeechQA, and tool calling, all while maintaining turn-based multimodal conversations. This means the model does not just understand the words being said; it can leverage emotion, intonation, and speaker-specific information to inform its responses.

This integration leads to a tangible leap in performance. In a comprehensive evaluation across 46 different benchmarks, A.X K2 Raon-Speech emerged as the top-performing open voice language model under 30B parameters for the Korean language, achieving a comprehensive score of 0.72. Its English performance was similarly strong, ranking third overall in the same category with a score of 0.75. The tension here is between model size and capability. By utilizing the MoE structure, Krafton has proven that a model with only 3.5B active parameters can outperform larger, dense models by specializing its internal experts for specific linguistic and auditory tasks.

When a model can handle tool calling and multimodal dialogue natively, the AI stops being a chatbot that can talk and starts becoming a voice agent that can act. The ability to process the nuance of a user's tone while simultaneously deciding which external tool to trigger represents a shift from passive response generation to active agency. The result is a system that reduces the cognitive load on the user and the computational load on the infrastructure.

This release signals a move toward a future where the boundary between text and speech is entirely erased in the weights of the model.