SK Telecom has passed the second-stage evaluation of the government’s “proprietary AI foundation model” project. The evaluation also reflected industrial applicability and ripple effects on South Korea’s AI ecosystem, leading to an assessment that it was recognised for high usability. SKT plans to expand its technology scope in the next evaluation to models with up to trillion-level parameters and into multimodal and AI agent areas.

SKT’s elite team passed the second-stage evaluation together with elite teams from Upstage and the LG AI Research Institute. The evaluation combined 40 points for benchmarks, 35 points for expert assessment and 25 points for user assessment.

An expert committee assessed that SKT secured top-tier performance among comparison models in mathematical reasoning and Korean-language areas, and demonstrated usability and practicality by deploying the model in large-scale commercial services. Industrial application cases were also reflected in the evaluation, including supplying A.X K1 to the defence sector, a proof-of-concept for a manufacturing-focused AI agent, and demonstrations in legal and tax fields.

◆From A.X K1 to K2… building technology for ultra-large models

SKT’s development of a proprietary foundation model gained momentum with A.Dot X K1 (A.X K1). A.X K1 is a from-scratch model in which SKT carried out everything in-house from model architecture design to pre-training using data, rather than additional training based on another company’s existing model.

It has 519 billion total parameters, but applies a mixture-of-experts structure in which only 33 billion are activated during actual token processing. This is designed to secure inference efficiency by expanding overall model capacity while avoiding computation of all parameters every time.

It also incorporated operational experience for infrastructure to train ultra-large models. SKT has worked with Nvidia since the 2021 build stage of the A100 GPU-based supercomputer “Titan” on data, infrastructure and the training environment. It also used Nvidia Nemotron datasets for training A.X K1. It applied the large-scale distributed training framework Megatron-LM and the data preparation and curation tool Nemo Curator.

A.X K2, used for the second-stage evaluation of the proprietary foundation model project, increased total parameters from 519 billion in K1 to 688 billion. It also applied a self-developed Sparse Gated Attention (SGA) structure. The approach selectively references highly relevant information rather than calculating all information equally when handling long contexts. SKT said this can improve accuracy and operational efficiency required in industrial settings.

Performance also improved versus K1. According to SKT, A.X K2’s average performance across 14 domestic and overseas benchmarks improved by 32.2 percentage points compared with K1. In particular, it improved by about 83.9 percentage points in long-context understanding and agent-related evaluations. In the second-stage evaluation results released by the Ministry of Science and ICT, its strong mathematical reasoning performance stood out. A.X K2 recorded a score meeting the gold-medal threshold in an evaluation using questions from the 2026 International Mathematical Olympiad. On the Math Arena AIME 2026 leaderboard, based on questions from the U.S. high school math contest AIME, it recorded a joint top score with 97.1 percent accuracy.

Its technology scope is also expanding from text-centred large language models to multimodal. SKT has built derived models centred on K2, including “A.X K2 VL Light-Preview,” which processes images and text together, and the speech model “A.X K2 ALM.” The vision model can be used for analysing industrial drawings and process manuals, and the speech model can be applied to call centres, meetings and on-site voice analysis.

◆Beyond model development, expanding application in industrial settings

Another pillar SKT emphasises in the project is application in industrial settings. It is applying A.X K2 to manufacturing-focused AI agents at steelmaker KG Steel and auto parts manufacturer Conec. In defence, it worked with the Ministry of National Defense to develop 3 quantised models based on A.X K1. The models are set to be used by units under the Army, Navy, Air Force and Marine Corps.

Separately, SKT’s elite team is running a consortium centred on 4 tracks: model and infrastructure, advanced research, data, and service expansion. The strategy is to connect model development and data building, the use of domestically made AI semiconductors, service verification and industrial application into a single system.

SKT plans to expand the scale of a successor to A.X K2, tentatively named “K3,” to trillion-level parameters. After scaling the model to 519B in the first stage and 688B in the second, it aims for further scale-up to compete on performance with global leading models. The government will also increase GPU support for the elite teams that advance to the next stage. The Ministry of Science and ICT plans to expand Nvidia B200 GPU support from about 768 units in the first half of this year to about 1,000 units in the second half.

SKT plans to widen the scope of applications for its proprietary models by leveraging existing service touchpoints such as A.Dot, A.Dot Biz and T Map, as well as industrial sites in manufacturing, defence and bio. Its goal is to settle proprietary models into actual services and industrial settings by combining vision, speech and agent technologies along with expanding model scale.

An SKT official said, “We will focus on strengthening perceived performance and agent execution capabilities,” and added, “We plan to continue advancing usability in daily life and industrial settings.”