SKT Releases Proprietary LLM 'A.X 3.1' as Open Source...Challenges Ultra-Large AI Model Development

Model with 34 Billion Parameters Released on Hugging Face
Two Independently Developed A.X 3.1 Models and Two Large-Scale Pretrained A.X 4.0 Models

SK Telecom announced on July 24 that it has released its self-developed large language model (LLM) to the open-source community and has also entered the Ministry of Science and ICT's 'Independent Artificial Intelligence (AI) Foundation Model Development Project.' Recently, SK Telecom was accepted as one of the 15 consortiums for the government-sponsored project, alongside Naver, LG AI Research Institute, and KT.

On this day, SKT unveiled its 'A.X (A.Dot X) 3.1' model, equipped with 34 billion parameters (34B), on the open-source platform Hugging Face. This model was built entirely in-house, from design to training, following a 'from scratch' approach. A.X 3.1 has improved code and mathematical performance, expanding its potential as an inference model, and achieved about 90% of the performance of major Korean benchmarks, despite being less than half the size of the previous A.X 4.0 (72B) model.

With this, SKT has now released a total of four models as open source: A.X 3.1 (standard and lightweight) and A.X 4.0 (large-scale pre-trained standard and lightweight). These models are available for both academic research and commercial use.

SKT is participating in this government project with a 'full-stack AI' strategy, covering everything from semiconductors to models, data, and services. The consortium includes leading companies in the AI value chain such as Krafton, FortyTwoDot, Rebellions, Riner, and SelectStar, as well as researchers from Seoul National University and KAIST. In addition, SK Group affiliates such as SK hynix, SK Innovation, and SK Broadband, along with K-AI Alliance companies including Moloco, CMES, Gauss Labs, and Scatter Lab, have submitted letters of intent to participate.

Through this consortium, SKT plans to develop an 'independent AI model representing Korea' and expand it into an 'omnimodal AI' capable of processing not only text but also images, voice, and video. The company is also conducting research on innovative new architectures that go beyond the current transformer structure and is taking on the challenge of developing ultra-large models for global technological independence.

Kim Taeyoon, head of SKT's foundation model division, said, "Through collaboration with leading companies in each field, we will deliver new achievements in the era of sovereign AI," adding, "We will present a comprehensive innovation model that takes into account the domestic AI ecosystem."

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