SKT Open-Sources Korean-Language Specialized A.Dot LLM, "Available for Enterprise Use"

"A Higher Score Than GPT-4o"
A.Dot X 4.0 Trained on Korean Data with Alibaba Qwen 2.5 as Its Base

SK Telecom announced on July 3 that it has open-sourced A.X 4.0, a Korean language-specialized large language model (LLM). On the same day, SKT released both the standard model and two lightweight models of A.X 4.0 through Hugging Face, a global open-source community.

SK Telecom announced on the 3rd that it has open-sourced A.X 4.0, a Korean language-specialized LLM. The photo shows SKT's self-built supercomputer 'Titan,' which conducted large-scale training for A.X 4.0. Photo by SKT

The newly released A.X 4.0 offers strengths such as efficient Korean language processing, a design that prioritizes data security, and the ability to operate in local environments. The model is based on China's Alibaba open-source LLM 'Qwen 2.5,' with additional Korean language data training.

According to SKT's internal testing, A.X 4.0 demonstrated approximately 33% higher token efficiency than GPT-4o when processing the same Korean sentences, indicating both high information processing capacity and cost-effectiveness.

In the Korean language proficiency benchmark (KMMLU), A.X 4.0 scored 78.3, surpassing GPT-4o's score of 72.5. In the Korean language and culture benchmark (CLIcK), it achieved a score of 83.5, again higher than GPT-4o's 80.2, proving its superior understanding of Korean culture.

Additionally, SKT plans to offer A.X 4.0 as an on-premise solution, allowing companies to install and use it directly on their internal servers rather than through the cloud, thereby alleviating concerns about data security. The standard model contains 7.2 billion parameters, while the lightweight models have 700 million parameters, enabling users to select the model that best fits their needs.

SKT has already applied A.X 4.0 to A.Dot call summarization since May, and plans to expand its use to various services within SK Telecom as well as across the SK Group in the future.

With the release of this model, companies can develop derivative models and utilize it in research fields. SKT aims to provide domestic companies with a new option to more easily leverage AI technology within their own environments.

Alongside the open-source release of the knowledge-based A.X 4.0 model, SKT is also preparing to announce an inference model. Within this month, SKT plans to unveil an inference model with enhanced mathematical problem-solving and code development capabilities, and will update the model to a level where it can simultaneously understand and process both images and text.

Furthermore, from a sovereign AI perspective, SKT is also developing models using the 'from scratch' approach applied to A.X 3.0, and will sequentially release follow-up models. 'From scratch' refers to building a model entirely from the very first stage.

Kim Jiwon, head of SK Telecom's AI Model Lab, stated, "We will continue to advance our technology to enhance SK Telecom's various services and to ensure that our Korean language-specialized LLM becomes the optimal model for the domestic business environment in the enterprise market."