SKT unveils Korean-specialized VLM on Hugging Face

SK telecom CI

SK telecom Co. announced Tuesday that it has open-sourced a new vision-language model (VLM) and document parsing technology for large language model (LLM) training on Hugging Face.

Both are based on the company’s AI platform A.X.

The newly released A.X 4.0 VL Light is a mid-sized model trained on large-scale multimodal Korean datasets. It is designed to interpret complex industrial data such as tables, graphs, and engineering schematics.

Built on the A.X 4.0 Light architecture, the model achieved an average score of 79.4 on Korean visual benchmarks, outperforming China’s Qwen 2.5-VL32B.

SK telecom also introduced the A.X Encoder, a high-speed data processing tool optimized for long documents.

Designed to support LLM training, the encoder delivers up to three times faster inference and double the training speed compared to existing models.

Operating with 149 million parameters, it achieved a natural language understanding benchmark score of 85.47, placing it at state-of-the-art (SOTA) levels globally.

With the release of these two technologies, SK telecom has unveiled a total of six models this month alone, signaling a strong push toward the national AI foundation model initiative.