SKT and Krafton release A.X-K2, a 688B MoE sovereign AI model
SK Telecom and Krafton released A.X-K2, a 688-billion-parameter Mixture-of-Experts model with 33 billion active parameters, as part of South Korea's sovereign AI effort. The package also includes specialized variants such as Raon-Speech, positioning a high-capability domestic open-weight stack for Korean-centric and multilingual use.
Key takeaway
National-scale foundation models are moving from policy slogans into shipped open-weight systems, giving builders a non-US alternative for Korean and multilingual workloads.
Context
A.X-K2 arrives as a sovereign AI deliverable from SK Telecom and Krafton rather than another incremental Western lab update. At 688B total parameters with 33B active in a Mixture-of-Experts design, it sits in the large open-weight class that local and enterprise teams actually evaluate for capability versus cost.
The release framing also stresses specialized variants such as Raon-Speech, which signals an ecosystem bet beyond a single chat checkpoint. For researchers and product teams, the practical question becomes licensing, eval quality, and serving economics versus US and Chinese peers already in wide circulation.
Numbers to know
- 688BTotal Mixture-of-Experts parameters in A.X-K2
- 33BActive parameters per forward pass