Rebellions, a South Korean AI semiconductor company, successfully operated SK Telecom's ultra-large AI model A.X K1 on domestic neural processing unit (NPU) servers, the company announced on the 23rd. Using its proprietary AI server 'Rebel Server,' Rebellions stably executed the model with over 500 billion parameters and demonstrated an AI agent service handling multiple user requests simultaneously. The achievement demonstrates the viability of operating large-scale AI services using solely domestic AI models and NPU infrastructure, positioning it as a practical example of sovereign AI capabilities. Rebellions is a consortium member under SK Telecom in the Ministry of Science and ICT's independent AI foundation model project, responsible for demonstrating AI infrastructure based on domestic NPUs.
Rebellions Optimizes Software for MoE Architecture and Parallel Processing
A.X K1 applies a Mixture of Experts (MoE) structure that selectively uses only necessary specialist models to enhance computational efficiency of large-scale models. Rebellions optimized dedicated software for the MoE model and applied technology to distribute various operations across NPUs for parallel processing. This enabled a single server to process multiple user requests in real time. The company stated that a single Rebel Server secures infrastructure performance comparable to global high-performance graphics processing unit (GPU) servers, reducing data center construction and operational burdens.
Rebel Server Demonstrates Real-Time Multi-User AI Agent Service
In the demonstration, Rebellions showcased an agent service where users could ask questions such as "pharmacies available on weekends" and receive results displayed on a map. The system processed responses in real time even in environments where multiple users simultaneously input different questions. The demonstration validated the server's capability to handle concurrent multi-user workloads with stable performance.
Rebellions Participates in Government AI Foundation Model Project
Park Sung-hyun, CEO of Rebellions, stated, "This is a practical case of sovereign AI showing that large-scale AI services can be operated using only domestic ultra-large models and domestic NPUs." Rebellions participates as a consortium member under SK Telecom in the Ministry of Science and ICT's independent AI foundation model project, where it is responsible for demonstrating AI infrastructure based on domestic NPUs.
FAQ
What did Rebellions achieve with SK Telecom's A.X K1 model?
Rebellions successfully operated SK Telecom's A.X K1 AI model, which has over 500 billion parameters, on its domestic NPU-based Rebel Server. The company demonstrated an AI agent service that handled multiple user requests simultaneously in real time.
How does Rebellions' Rebel Server handle large AI models efficiently?
Rebel Server optimizes software for the Mixture of Experts (MoE) architecture used in A.X K1, distributing various operations across NPUs for parallel processing. This enables a single server to achieve infrastructure performance comparable to global high-performance GPU servers while processing multiple user requests in real time.
