[Edaily Reporter Yun Junghoon ]SKTelecom(017670)’s proprietary artificial intelligence (AI) foundation model, “A.X K2,” is proving its performance is on par with the world’s top open-source AI models, thereby strengthening its presence in the domestic AI competition.

Strengths in Math and Korean... Outperforms Models Like DeepSeek and Kimi

According to industry sources on the 29th, SKT unveiled “A.X K2,” a super-large AI model with 688 billion (688B) parameters, via Hugging Face that day. While expanding in scale compared to its predecessor, A.X K1 (519 billion parameters), it has significantly improved its capabilities in mathematical reasoning, solving scientific problems, understanding Korean, and processing long texts.

A.X K2 achieved performance on par with its peers across 14 major benchmarks when compared to leading global open-source models released within the past six months, including Alibaba’s Qwen 3.5, DeepSeek V4 Flash, GLM 5.1, and Kimi K2.6.

In particular, it outperformed competing models in the Korean language and mathematics domains. It scored 97.1 on the AIME26 mathematics assessment, 80.5 on the KMMLU-Pro Korean language comprehension test, and 91.6 on the CLIcK test—the highest scores among the models compared. It also demonstrated top-tier performance on the τ²-Bench (Telecom), which evaluates specialized question-answering capabilities in the telecommunications field, scoring 98 points.

Real-world performance has also improved. A.X K2 scored 35 out of a possible 42 points on past International Mathematical Olympiad (IMO) 2025 problems and solved all eight questions from the second round of the 2026 Korean Mathematical Olympiad (KMO). It maintained its ability to accurately locate specific information even when processing long-form documents containing approximately 256,000 tokens, thereby increasing its potential for use in enterprise AI services.

This achievement is particularly noteworthy because it demonstrates the ability to overcome limited computing resources through software technology.

Trained for 70 days using 512 B200 GPUs... Benchmark performance improved through proprietary technology

Recently, Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon emphasized the need to expand the national AI infrastructure, stating, “With the current level of support—735 NVIDIA B200 GPUs—there are limitations to developing world-class frontier AI models.”

In this context, SKT trained 8.5 trillion tokens over approximately 70 days using 512 NVIDIA B200 GPUs. Although the number of training tokens was lower than that of its predecessor, the A.X K1, average performance on domestic and international benchmarks improved by 32.2 percentage points.

Behind this performance improvement lies SKT’s proprietary AI architecture. Its flagship technology, “SGA (Sparse Gated Attention),” is designed to selectively reference only the necessary information when processing long documents, thereby increasing computational efficiency. In addition, the company applied “Gated Norm” technology—which reduces weight instability during the training process—to enhance the stability of model training. Furthermore, by training the model in FP8 format from the outset—without undergoing a separate lightweighting process—the company reduced storage requirements and inference costs by nearly half.

Currently, four teams—including SKT, LG Corp., Upstage, and Motif Technologies—are undergoing the second-stage evaluation for the government’s Independent AI Foundation Model Project (Dokpamo). From the 8th to the 11th of next month, a real-world evaluation will be conducted with a panel of 200 citizens. In addition to existing benchmark scores, actual user experience—including the accuracy, usability, and satisfaction of responses—is expected to be a key criterion in the final evaluation. Furthermore, plans are in place to closely examine, through expert evaluations and other means, the models’ scalability in real-world industrial settings.

Each company’s strategy is also distinct. SKT has secured both performance and efficiency by applying a Model of Experts (MoE) architecture, which uses only 33 billion of the total 688 billion parameters for actual inference. LG Corp. AI Research is emphasizing multimodal scalability, Upstage is focusing on practical AI that can run even in small-scale GPU environments, and Motif Technologies is highlighting resource efficiency based on its proprietary architecture.

SKT is expanding the application of A.X K2 to industrial settings such as manufacturing, defense, and biotechnology. It is conducting pilot projects for manufacturing Agent AI with KG DONGBUSTEEL and Konec, and is collaborating with the Ministry of National Defense to develop quantized models for deployment in military operational environments. The company is also expanding the application of AI to A.dot and SK hynix(000660) for business-use AI. Additionally, it is exploring ways to utilize A.X K2 ALM—which supports real-time voice recognition and analysis—for customer service center calls.

Kim Tae-yoon, Head of Foundation Models at SKTelecom, stated, “We will continue to apply and develop A.X K2 across various industries to contribute to strengthening national competitiveness.”