South Korea Just Got Two New 700B+ Open-Source AI Models in Two Days

South Korea released two frontier-scale open-source AI models within 48 hours of each other. SK Telecom published A.X K2, a 688-billion-parameter model, on Hugging Face on July 29, 2026; LG AI Research followed on July 31 with K-EXAONE 2.0, a 750-billion-parameter model. Both are competing for the same prize: Korea’s National AI Foundation Model project, a government initiative to prove the country can build frontier-class AI entirely with domestic technology.

Quick facts

• K-EXAONE 2.0 (LG AI Research): 750 billion total parameters, 37 billion active per token, released July 31, 2026 under Apache 2.0.
• A.X K2 (SK Telecom): 688 billion parameters, released July 29, 2026, also Apache 2.0 licensed.
• Both use a Mixture-of-Experts architecture with a 262,144-token context window — nearly identical technical blueprints from competing teams.
• K-EXAONE 2.0 scored 70.1 average across 24 benchmarks, up from 63.3 for its 236-billion-parameter predecessor — a jump of over 10%, with a roughly 30% improvement specifically on coding and agentic-coding benchmarks.
• On long-context comprehension (OpenAI-MRCR), K-EXAONE 2.0 scored 94.4, ahead of the 71.5 LG reports for Zhipu AI’s GLM-5.1 on the same test.

Why two labs shipped almost the same thing at almost the same time

This isn’t a coincidence of timing so much as a shared deadline. Both companies are competing teams inside South Korea’s Independent (Sovereign) AI Foundation Model Project, run by the Ministry of Science and ICT, which is heading into a second-phase evaluation round. According to TechTimes’ reporting, a third major team — Motif Technologies — has not yet released a comparable public model ahead of the evaluation, leaving LG and SK Telecom as the two clearest public data points so far.

Both companies made the same licensing bet, too: full Apache 2.0, the same permissive license Meta uses for Llama, allowing any company anywhere to download, modify, and deploy either model commercially with no royalty and no obligation to release their changes. For LG specifically, that’s a real shift — The Elec reports that earlier EXAONE releases used more restrictive licensing, making K-EXAONE 2.0 the first fully commercially permissive release in the series.

What K-EXAONE 2.0 actually improved

LG’s model more than tripled in size from its 236-billion-parameter predecessor, and the performance gains tracked with that: a 10%+ jump in average benchmark score across 24 tests spanning nine categories — knowledge, math, coding, agentic tasks, instruction following, long-context understanding, multilingual performance, and safety — plus a roughly 30% improvement specifically on coding and agentic-coding evaluations. LG also added Multi-Token Prediction and a technology it calls DSpark, which the company says makes text generation three to five times faster during inference. Language support expanded to 10 languages, up from a narrower Korean-and-English focus.

LG’s own technical report is notably candid that the model doesn’t win every comparison — it acknowledges trailing some competitors in specific areas, even as it leads on long-context tasks like OpenAI-MRCR and the Korean-language Ko-LongBench benchmark, where LG reports K-EXAONE 2.0 beat China’s GLM-5.1 by more than 10% on average across the three long-context tests it compared.

What this means if you’re choosing an open-weight model

For teams evaluating open-source models to self-host or fine-tune, both releases are immediately usable under fully commercial terms, and both now sit in the same size and architecture class as the leading Chinese open-weight releases they’re explicitly benchmarked against. The practical differentiators right now are strongest multilingual coverage (K-EXAONE 2.0’s 10 languages) versus SK Telecom’s own positioning for A.X K2, and long-context performance, where LG’s published numbers currently lead. Neither model has had significant independent, third-party benchmark verification yet — the figures above come from each company’s own technical reporting.

Side by side: K-EXAONE 2.0 vs. A.X K2

• Developer: K-EXAONE 2.0 — LG AI Research. A.X K2 — SK Telecom.
• Release date: K-EXAONE 2.0 — July 31, 2026. A.X K2 — July 29, 2026.
• Total parameters: K-EXAONE 2.0 — 750 billion. A.X K2 — 688 billion.
• Architecture: both use a Mixture-of-Experts design with a 262,144-token context window — K-EXAONE 2.0 activates roughly 37 billion parameters per token, selecting 8 of 256 specialized expert modules for each generated token.
• License: both Apache 2.0, fully commercial, no restrictions.
• Predecessor size: K-EXAONE 2.0 more than tripled its 236-billion-parameter predecessor; SK Telecom has not published an equivalent first-generation comparison for A.X K2 in the reporting reviewed here.

The architectural convergence is itself notable: two separate Korean teams, working independently under the same government program, landed on nearly identical technical choices — MoE routing, the same context window length, the same licensing model. That’s less a coincidence than a signal about where the current competitive frontier for mid-size sovereign AI models actually sits right now.

Key takeaway

Two competing Korean teams just put frontier-scale, fully commercial open-weight models into the same window as the leading Chinese open releases, and did it under deliberate government pressure to prove Korea can build this domestically. Whichever model wins the government evaluation, both are already downloadable and usable today — worth a real evaluation against your own workload rather than taking either company’s benchmark numbers at face value.