SKT launches A.X-K2, South Korea's 21B-parameter sovereign AI model
A massive MoE model with 3B active parameters, backed by $360M national funding.
SK Telecom (SKT) has released A.X-K2, a Mixture-of-Experts (MoE) foundation model boasting 21 billion total parameters with 3 billion active per token. The model is publicly available on Hugging Face alongside A.X-K2-ALM for alignment tuning. In collaboration with KRAFTON, SKT also released A.X-K2-Raon-Speech-21B-A3B, a speech-focused variant. These releases are central to South Korea's Sovereign AI Foundation Model Project, a national initiative to reduce dependence on foreign AI.
The project, running through 2027, has a government budget of ₩530 billion (approximately $360 million) distributed among multiple companies. Initially five firms received funding: Upstage, SKT, LG AI Research, Naver Cloud, and NC AI. After the first evaluation in December 2025, Naver Cloud and NC AI were dropped, and Motif Technologies was added in February 2026. The second evaluation is scheduled for August 2026. The A.X-K2 release positions SKT as a leading contender in South Korea's bid for AI sovereignty.
- A.X-K2 is a MoE model with 21B total parameters (3B active), plus alignment and speech variants released.
- Part of South Korea's Sovereign AI Project with ₩530B ($360M) government investment through 2027.
- Upstage, SKT, LG AI Research, and Motif Technologies remain funded; Naver Cloud and NC AI were cut after December 2025 evaluation.
Why It Matters
South Korea secures AI independence with a domestically built foundation model rivaling global leaders.