← Back to feed

South Korea’s government-backed sovereign AI competition reached its most competitive moment this week. SK Telecom published A.X K2, a 688-billion-parameter open-weight model, on Hugging Face on July 29. LG AI Research followed two days later with K-EXAONE 2.0 at 750 billion parameters. Both use MoE architectures, both carry Apache 2.0 licenses, and both were built to influence the Ministry of Science and ICT’s second-phase evaluation running August 8-11. That evaluation narrows the field from four teams to three.

The Models

A.X K2 (SK Telecom) — 688B total, 33B active, trained from scratch in FP8 precision over 8.5 trillion tokens across 70 days on 512 NVIDIA B200 GPUs. SK Telecom built its own Sparse Gated Attention architecture, which attends only to the top 2,048 most relevant positions per query rather than the full context — delivering a 262,144-token window at lower memory cost. The FP8 weights ship as-released, removing the typical quantization step for deployment.

K-EXAONE 2.0 (LG AI Research) — 750B total, 37B active, expanded from a 236B predecessor. The architecture combines global attention layers (full document) with sliding-window local attention, layered on top of a 256-expert MoE. A 262,144-token context window with the same hybrid routing as the original K-EXAONE family.

Benchmark Divergence

The two models have non-overlapping strengths, which is the most useful detail ahead of the government evaluation.

A.X K2 leads on:

• KMMLU-Pro (Korean knowledge): 80.5 — top in its comparison set
• CLIcK (Korean language): 91.6 — top in its comparison set
• AIME26 (mathematics): 97.1
• IMO 2025: 35 of 42 problems solved, above the gold-medal threshold
• τ²-Bench Telecom: 98.0
• Apex (hard math): 45.8 vs 28.1 for runner-up DeepSeek-V4 Flash

A.X K2 trails significantly on:

• BrowseComp: 9.3 — last in its comparison set, against GLM-5.1 at 29.1 and Qwen3.5 at 26.9. SK Telecom’s own model card attributes the gap to “limited agentic RL during post-training” and advises users to evaluate the model before deploying in long tool-use workflows.

K-EXAONE 2.0 leads on:

• OpenAI-MRCR (long-context retrieval): 94.4 — above Qwen3.5 (93.0) and DeepSeek V4 Pro Max (92.9)
• Ko-LongBench (Korean long-context): 89.6 — above GLM-5.1 (83.6)
• Tau3-Bench Banking: 14.2 — above GLM-5.1 (11.5) and Qwen3.5 (13.4)
• 24-benchmark average: 70.1, up from 63.3 for the predecessor (over 10% gain)
• Agentic coding: approximately 30% improvement over K-EXAONE 1.0

LG AI Research scored highest in the first government evaluation round with 90.2 out of 100, on a three-part assessment covering benchmarks (40 points), expert panel review (35 points), and user feedback (25 points).

The Competition Structure

The Ministry of Science and ICT started this project in August 2025 with five teams. Two were eliminated in January 2026: Naver Cloud, for building on frozen Alibaba Qwen encoder weights (violating the domestic-technology requirement), and NC AI. A startup consortium — Motif Technologies, KAIST, and Trillion Labs — joined in February 2026 to fill the gap, bringing the current field to four.

The four remaining teams:

• LG AI Research — K-EXAONE 2.0, 750B MoE (Phase 1 leader: 90.2/100)
• SK Telecom — A.X K2, 688B MoE
• Upstage — Solar Open 2, 250B, focus on practical deployment in limited GPU environments
• Motif Technologies — Proprietary architecture, 300B+ total, 13B active, startup-built from scratch

The August 8-11 evaluation introduces 200 citizen evaluators testing real-world usability alongside benchmark scores. The final two-team selection happens by year-end; those models will serve as the technical backbone for South Korea’s national AI chatbot, covering 51 million citizens.

The $5.7 billion government investment through the National Growth Fund, approved in May 2026, provides the infrastructure context. Korea’s Ministry of Science and ICT has a 2026 AI budget of roughly $3.5 billion.

The Sovereignty Question

The “sovereign” label warrants close reading. A.X K2’s training data is 72.7% English and 8.3% code, with Korean at 15.4%. The English corpora rely on NVIDIA’s Nemotron-CC-v2.1 and FineWeb2 — public international datasets, not domestic data. What is genuinely domestic is SK Telecom’s 1.37-trillion-token proprietary Korean corpus and the SGA architecture. LG AI Research’s model carries similar caveats.

What both labs actually achieved is weight sovereignty: the ability to own, modify, and redistribute model weights domestically without API dependency. For a government that watched Naver Cloud disqualified for using frozen foreign weights, that distinction is the point.