SK Telecom releases A.X K2, a 688B open-weight model built for sovereign AI

Korea just dropped one of the largest open-weight AI models in the world.

On July 29, 2026, SK Telecom unveiled A.X K2, a 688 billion parameter mixture-of-experts foundation model, and released the weights on Hugging Face for anyone to download.

The numbers over its predecessor

Against A.X K1, the 519 billion parameter model it replaces, K2 posts substantial gains:

• +32.2 points average across 14 domestic and international benchmarks
• +83.9 points on long-context comprehension and agent evaluations

The agent and long-context jump is the more telling figure. Those are the capabilities that decide whether a model can run real workflows rather than answer isolated questions, and they are where the current generation of frontier models competes hardest.

Sovereign AI in action

What makes this bigger than a benchmark story is the strategy behind it. A.X K2 is sovereign AI in action: SK Telecom is positioning the model for Korea’s critical sectors, manufacturing, defense and biotech, built and controlled at home instead of rented from US labs.

That framing has been spreading through national AI strategies all year. Nations increasingly want their own frontier models, not API keys to someone else’s. A domestically built, domestically controlled model at 688B parameters is a statement that Korea intends to be a producer in that world, not just a customer.

The open-weight arms race continues

The release also keeps pressure on the open-weight front. It lands a week after Kimi K3’s 2.8 trillion parameter release, continuing a pattern where frontier-class models keep arriving on Hugging Face for anyone to download.

Every release at this scale resets the floor for what closed labs have to justify charging for. And every country that ships one moves the sovereign AI conversation from policy papers to working code. The next question is simply which country ships the next one.