Released 688B Model on Hugging Face
Enhanced Reasoning in Math, Science, and Korean Knowledge
Expansion into Manufacturing, Defense, and Bio Industries
LG AI Research Boosts Benchmark Performance
Upstage and Motif Technologies Advance Model Sophistication

SK Telecom Unveils Second ‘A.X K2’ Model for Korean-only Foundation Project... Enhanced Reasoning and Long-form Proficiency Set It Apart (Comprehensive) View original image

As the government's second evaluation for the 'Independent Artificial Intelligence (AI) Foundation Model (Korean-only Foundation Model)' project approaches in August, four elite teams are upgrading their models by significantly increasing the number of parameters and improving reasoning abilities.


On July 29, SK Telecom announced plans to release its enhanced AI foundation model 'A.X K2' on the open-source community Hugging Face and to expand its applications into manufacturing, defense, bio, and other industrial fields. 'A.X K2' has expanded its parameters from 519 billion (519B) in the previous 'A.X K1' model to 688 billion (688B), and strengthened mathematical and scientific reasoning, as well as Korean knowledge and long-form reasoning abilities.


Compared to the previous model, A.X K2's average performance across 14 domestic and international benchmarks improved by 32.2 percentage points. The model showed a particularly significant improvement in long-form understanding and agent-related evaluations, with an increase of about 83.9 percentage points. SK Telecom explained that it has secured performance equivalent to recently released open-source models from abroad, including Qwen, DeepSeek, and Kimi.


In the manufacturing sector, practical on-site know-how has been embedded into AI. SK Telecom is applying A.X K2 to manufacturing-specialized AI agents in partnership with KG Steel, a steel manufacturer, and Konec, an auto parts manufacturer. In the second half of the year, a demo version of the manufacturing-specialized AI agent will be piloted at KG Steel's Dangjin plant cold rolling line—producing galvanized steel sheets—and at Konec's casting and processing lines for real-world verification.


Additionally, in collaboration with the Ministry of National Defense, SK Telecom has developed three quantized models based on A.X K1. The Ministry and SK Telecom plan to deploy these models in units under the Army, Navy, Air Force, and Marine Corps. SK Telecom is also utilizing AI in partnership with SK Biopharmaceuticals for the development of targeted therapies for intractable cancers.


SK Telecom Plans to Expand the Next 'A.X K2' Model to Trillion-Parameter Scale


SK Telecom will also participate in the 'AI for All' project promoted by the Ministry of Science and ICT, and aims to expand its next A.X K2 model to a trillion-parameter scale to secure global competitiveness.


Furthermore, SK Telecom has built a derivative lineup around the newly released A.X K2, including vision-language (VL) models based on vision encoders and speech models based on audio encoders and decoders, to address diverse industrial demands. On the same day, Krafton, a member of the SK Telecom consortium, released its independently developed model, 'A.X K2 Raon-Speech.' A.X K2 Raon-Speech features a small language model based on SK Telecom's A.X K2, enhanced by Krafton's own speech encoder and codec, trained to understand and generate speech.


LG AI Research is differentiating itself ahead of the second evaluation by enhancing the performance of 'K-Exaone.' By scaling up from the previous model, the institute is focusing on sophistication to compete with globally advanced frontier large language models (LLMs), and has improved benchmark performance across multiple areas including knowledge, reasoning, instruction-following, and long-context processing. Capabilities in agentic tool utilization and agentic coding—areas with rapidly increasing real-world applications—are also being strengthened. An LG AI Research official stated, "We are preparing for public and real usability evaluations."



Upstage revealed the real-world usability potential of its model for the second evaluation, 'Solar Open2.' The number of parameters has been expanded to 250 billion (250B), with a focus on agent-based practical applications. Motif Technologies has distinguished itself by developing its own proprietary architecture entirely in-house, rather than modifying overseas open-source models. Lim Junghwan, CEO of Motif Technologies, said, "By applying the MoE (Mixture of Experts) structure, we maintain a total of over 300 billion parameters, but only about 13 billion (13B)—approximately 4%—are selectively activated during the reasoning process, enhancing computational efficiency." He added, "We also prioritize keeping the number of active parameters in MoE models from becoming too large, so as not to slow down reasoning speed."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing