SK Telecom debuts Korea's first 519 billion-scale AI model

SK Telecom has unveiled A.X K1, a hyperscale AI model with 519 billion parameters, marking Korea’s first entry at this scale.

Developed by a consortium including SK Telecom, Krafton, 42dot, Rebellions, Seoul National University, and KAIST, the model is positioned as a foundational platform for Korea’s national AI infrastructure.

A.X K1 is designed to transfer knowledge to smaller AI models and will be available through open-source channels and APIs.

The consortium said it will be integrated into services with over 10 million users and industry applications, while more than 20 institutions will join real-world testing and validation.

🔗 Source: SK Telecom

🧠 Food for thought

Implications, context, and why it matters.

Parameter count hype needs real-world validation benchmarks

A.X K1 lists 519 billion parameters. That number means little without independent testing on benchmarks such as KMMLU-Pro 1.
KMMLU-Pro tests 2,822 problems from Korea’s 2024 National Professional Licensure exams across 14 professions, giving domain coverage beyond generic language tasks 1.
If A.X K1 uses a mixture-of-experts (MoE) setup where only some parameters fire per input, results will differ from a dense model that activates all parameters. Scores on Codeforces (a competitive programming platform) and the American Invitational Mathematics Examination (AIME) are needed to judge parity with Kimi k1.5, a reasoning model that claims state-of-the-art results via reinforcement learning scaling 2.
The A.X K1 consortium plans to release parts of the training data. Details on API endpoints, hosting needs, and hardware support are still missing.

Open APIs create immediate market for Korean enterprise AI infrastructure

Cloud firms and AI infra shops can use A.X K1 API access to offer private hosting for Korean companies. Many need data sovereignty plus local rules compliance.
Vendors can ship fine-tuning tools for Korean manufacturers or gaming firms that adapt A.X K1 to specific workloads. Larod provides unified machine learning (ML) interfaces for hardware acceleration 3.
DevOps tooling firms can build deployment frameworks that connect A.X K1 with existing Korean enterprise systems. That work patches integration gaps that slow AI rollouts.
The consortium includes Rebellions, a semiconductor company. Hardware vendors can pitch inference accelerators to boost A.X K1 performance if model details expose compute bottlenecks.