"The most significant change in A.X K2 is the shift from a 'reactive AI' to a 'proactive AI' that can plan and execute tasks on its own," said Kim Tae-yoon, head of SK Telecom's (SKT) foundation model division, on August 4 in a statement released through the company's newsroom. He noted that the model's reasoning capabilities have been greatly enhanced to align with the era of agentic AI, expanding the overall model size to 688 billion parameters while maintaining an active parameter count of 33 billion during inference to improve efficiency.
SKT emphasized its focus on performance improvements tailored for AI agent services. By applying its proprietary Sparse Gated Attention (SGA) structure, the model can reliably handle long contexts, demonstrating superior performance in throughput and latency compared to its predecessor, A.X K1, especially as input length increases. In environments processing documents over 300 pages long, throughput improved by 67.7% compared to the previous version.
In addition to text-based models, SKT unveiled derivative models that support visual and audio capabilities. The vision model, A.X K2 VL Light-Preview, integrates a self-developed vision encoder, A.X VE, enabling analysis of manufacturing blueprints, documents, and charts. In the audio domain, SKT introduced its audio language model, A.X K2 ALM, along with A.X K2 Raon-Speech, developed in collaboration with Krafton. Kim stated, "The industrial field does not operate solely on text; AI that encompasses various modalities is necessary for practical applications."
SKT is already applying A.X K2 in real industrial settings. It is conducting a pilot project for a manufacturing-specific AI agent with KG Steel and automotive parts company Conec, developing a system that learns from production line defects to analyze causes and propose solutions. A lightweight model is also being utilized in SK Hynix's internal AI tool, LLM Chat, for semiconductor knowledge searches and information organization.
Looking ahead, SKT plans to release its proprietary AI model under a free license to support commercial use by developers and startups, and will facilitate service commercialization through an accelerating program. Future models are expected to expand to a trillion-parameter scale and continue research into agentic reinforcement learning, cybersecurity capabilities, and next-generation AI architecture.
Kim concluded, "A.X K2 demonstrates the potential for a model developed with domestic technology to reach the highest levels in mathematics and Korean language fields. We aim to translate this technology into tangible value that customers and industries can experience."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.