SK Telecom announced on the 3rd that it has released the Korean language specialized large language model (LLM) 'adot A.X 4.0' as open source.
On this day, SK Telecom released the standard model and lightweight model of A.X 4.0 through the open-source community Hugging Face. The standard model has 72 billion (720억개) parameters, while the lightweight model has 7 billion (70억개) parameters. The company noted that A.X 4.0 boasts one of the highest Korean processing efficiencies among existing LLMs and emphasized its strengths in the design considering data security and the increased possibility of operation in a local environment.
adot A.X 4.0 was developed by adding Korean data to the LLM 'Qwen 2.5', which was released as open-source by the Chinese corporation Alibaba. SK Telecom said it completed all stages of large-scale training (CPT · Continual Pre-Training) of A.X 4.0 using its own data without external integration, ensuring 'data sovereignty.'
Although the basis of A.X 4.0 was derived from Alibaba, the tokenizer, a tool for analyzing sentence structure and partitioning into tokens, was designed and implemented by SK Telecom itself. The company explained that this enables a high level of Korean processing capability. According to SK Telecom’s internal test results, when the same Korean sentences were input, A.X 4.0 recorded about 33% higher token efficiency than OpenAI's 'GPT-4o'.
A.X 4.0 scored 78.3 points on 'KMMLU', a representative Korean language ability assessment benchmark, outperforming OpenAI's latest AI model GPT-4o (72.5 points). It also received 83.5 points on CLIcK, a Korean language and cultural benchmark, scoring higher than GPT-4o (80.2 points), illustrating its understanding of Korean culture.
SK Telecom provides A.X 4.0 in an on-premise (internally constructed) manner. It can be directly installed on corporate internal servers, offering strengths in data security.
SK Telecom stated, 'We have already applied A.X 4.0 to the currency summary of adot in May, and it is being successfully utilized,' adding that it plans to apply it to various services within the SK Group in the future. It continued, 'With A.X 4.0, corporations can develop derivative models and utilize it in research fields,' and expressed its intention to provide a new option that allows domestic corporations to more easily utilize AI technology in their own environments.
SK Telecom is set to announce an inference model simultaneously with the open-source release of the A.X 4.0 knowledge model. This month, SK Telecom plans to release a reasoning model with enhanced capabilities for solving mathematical problems and code development and aims to update the model to a level that can simultaneously understand and process images and text. It is also proceeding with development using the 'Sovereign AI' perspective, employing a from-scratch method applied to A.X 3.0.
Kim Ji-won, head of SK Telecom's AI model lab, noted, 'We plan to promote continuous technology development to enhance various services of SK Telecom, making it an LLM specialized for Korean language that can be optimized for the domestic business environment.'