SK Telecom has unveiled its proprietary AI Foundation model with 688 billion parameters and is stepping forward to support AI transformation that spans industrial sites and everyday life. This unveiling demonstrates that SKT's AI model lineup has transitioned into a full-fledged multimodal system, as it not only significantly boosted performance compared to the previous model but also introduced vision and voice derivative models.

SKT announced on the 29th that it has released its proprietary AI foundation model, AX K2, to the open source community Hugging Face.

AX K2 has 688 billion (688B) parameters, increasing the model size compared to its predecessor, AX K1 (519B), thereby enhancing mathematical and scientific reasoning and Korean long text comprehension capabilities.

The average performance of 14 domestic and international benchmarks improved by 32.2%p compared to AX K1, and in long text comprehension and agent-related items, it improved by 83.9%p.

It is designed to selectively reference only highly relevant information when processing long contexts by applying SKT's proprietary SGA (Sparse Gated Attention) structure.

It is reported to have recorded benchmark results equivalent to those of overseas models such as Qwen3.5-397B-A17B, DeepSeek-V4-Flash, GLM-5.1, and Kimi-K2.6.

SKT is applying AX K2 to various fields such as manufacturing, defense, and bio.

In the manufacturing sector, we plan to establish manufacturing-specialized AI agents at KG Steel and KONEX, and conduct field trials in the second half of the year at KG Steel's Dangjin plant's cold rolling line and KONEX's casting and processing processes.

In the defense sector, three AX K1-based quantization models were developed in collaboration with the Ministry of National Defense, and there are plans to utilize them in units under the Army, Navy, Air Force, and Marine Corps.

In the bio sector, SK Biopharmaceuticals applied AI to the development of targeted therapies for intractable cancers. As a result, it was announced that the initial research period for new drug development, which typically takes 1 to 2 years, has been shortened to 5 months.

In office environments, AX K2 is integrated into A.Biz to provide newsletter creation and report generation functions, and a lightweight model is also being supplied to SK Hynix's in-house business AI tools.

SKT also unveiled three derivative models based on the AX K2.

△AX K2 VL Light-Preview (21B): A vision language model based on a vision encoder that understands images and text simultaneously and can be used for analyzing industrial drawings and legal documents. △AX K2 ALM (22B): A speech language model with proprietary voice encoders and decoders applied, currently under review for use in customer center voice calls. △AX K2 Raon-Speech (21B): A speech language model independently developed by Krafton, which is planned to be applied to AI voice interactions within games in the future.

SKT plans to secure global competitiveness by expanding the successor model to a parameter scale in the trillions of won.

"We have enhanced various functions so that it can be widely used in everything from individuals' daily lives to office environments and manufacturing sites," said Kim Tae-yoon, head of SK Telecom Foundation Model. "We will continue to apply and develop the AX K2 across various industrial sectors to contribute to strengthening national competitiveness."