South Korea's two largest sovereign AI competitors released rival foundation models within 48 hours of each other this week, marking the most direct head-to-head moment yet in a government-backed race that will cut its field from four teams to three when evaluators convene August 8 through 11. SK Telecom published A.X K2, a 688-billion-parameter open-weight model, on July 29, while LG AI Research followed with K-EXAONE 2.0, a 750-billion-parameter model, on July 31. Both models use nearly identical architectural blueprints, including Mixture of Experts (MoE) design, 262,144-token context windows (roughly equivalent to processing several full-length novels in a single pass), and Apache 2.0 licensing that allows any company to use or modify them commercially without restriction.

What Do These Massive Parameter Counts Actually Mean?

The headline numbers sound enormous, but they describe something different than most people assume. A model with 750 billion "total parameters" in a Mixture of Experts design does not engage all 750 billion during any single response. Instead, the model divides its neural network into 256 specialized processing modules, called "experts," and dynamically routes each incoming text token through only eight of them. For K-EXAONE 2.0, the result is that only approximately 37 billion parameters are active during inference; for A.X K2, approximately 33 billion are active per token. This is a design choice, not a limitation. The core idea is that a model can accumulate vast breadth of knowledge across 750 billion parameters while operating at the inference cost of a much smaller dense model, giving the model a wider range of specialized representations to draw on without proportional computing burden.

The two models achieve their massive context windows through different mechanisms. K-EXAONE 2.0 layers a hybrid attention mechanism on top of the MoE design, alternating between global attention layers that see the full document and local sliding-window attention layers that process nearby text in chunks. A.X K2 reaches the same context length through SK Telecom's proprietary Sparse Gated Attention, which keeps only the top 2,048 most relevant positions per query rather than attending across all prior tokens.

How Did Each Team Build Their Model Differently?

The most technically significant difference between the two releases is how they were constructed. K-EXAONE 2.0 was produced through what LG calls "upcycling": the team took the existing 236-billion-parameter K-EXAONE model, LG's Phase 1 submission to the government competition, and expanded it in both depth and width before retraining. The 750-billion-parameter result inherits architecture and some learned representations from K-EXAONE 1.0, then undergoes continual pretraining, difficulty-focused mid-training, and post-training on the expanded network. LG found that applying a clamping operation after activation branches in each layer prevented the exploding activations that otherwise destabilize training in deeper networks.

A.X K2 took a fundamentally different path. SK Telecom trained the model natively in FP8 precision, a lower-precision floating-point format than the BF16 most frontier labs use, for both the forward and backward passes throughout training, across approximately 8.2 trillion tokens. This remains uncommon at this model scale. The released model weights are also shipped in block-scaled FP8, meaning a user running A.X K2 can deploy it at lower memory cost without a separate quantization step. SK Telecom applied its proprietary Sparse Gated Attention throughout training, which it credits for suppressing the activation-range problems that typically plague FP8 training at large scale.

Where Do These Models Excel and Struggle?

On publicly reported benchmarks, the two models have complementary strengths. K-EXAONE 2.0 posted the strongest long-context retrieval scores in LG's comparison set: 94.4 on OpenAI-MRCR, above Qwen3.5 (93.0) and DeepSeek V4 Pro Max (92.9), and 89.6 on Ko-LongBench, the Korean-language long-context benchmark, above GLM-5.1 (83.6). Across its 24-benchmark evaluation suite, K-EXAONE 2.0 averaged 70.1, compared with 63.3 for its predecessor, an improvement of over 10 percent. Agentic coding benchmarks improved by roughly 30 percent over K-EXAONE 1.0.

A.X K2 performed strongly on mathematics: 97.1 on AIME26, 35 of 42 problems on the International Mathematical Olympiad 2025 (above the gold-medal threshold), and 80.5 on KMMLU-Pro. SK Telecom also reported an average 32.2-percentage-point gain over A.X K1 across 14 domestic and international benchmarks.

However, one number in A.X K2's model card reveals an important weakness. On BrowseComp, a benchmark measuring a model's ability to answer difficult questions through open-ended web research, A.X K2 scored 9.3, finishing last in SK Telecom's comparison set, against 29.1 for GLM-5.1 and 26.9 for Qwen3.5. SK Telecom's own model card disclosed this result and attributed it to "limited agentic RL during post-training," recommending users "evaluate long-horizon tool-use workloads before relying on them." A model that excels at structured mathematical reasoning while struggling at open-ended web research reveals where its post-training effort concentrated, precisely the kind of capability gap that government and enterprise buyers need to evaluate before deploying these models in agentic workflows.

How Did South Korea's Sovereign AI Race Get to This Point?

The competitive landscape around these releases is not the one most international coverage has described. When South Korea's Ministry of Science and ICT launched its Sovereign AI Foundation Model project in August 2025, five consortia were selected: LG AI Research, SK Telecom, Upstage, Naver Cloud, and NC AI. The government's plan called for one team to be eliminated after the first evaluation in December 2025. On January 15, 2026, MSIT announced that Naver Cloud and NC AI had failed to advance. Naver Cloud, previously considered a frontrunner given its HyperCLOVA X heritage and massive Korean-language data advantage, was cut because its model submission used frozen encoder weights from Alibaba's Qwen model for video and audio components, which MSIT determined did not meet the project's requirements.

This elimination left three teams competing for the second phase: LG AI Research, SK Telecom, and Upstage. The government's second evaluation, scheduled for August 8 through 11, will cut the field from three teams to two, making the timing of these two model releases particularly significant. Both teams are essentially making their case to evaluators with competing technical approaches and benchmark profiles.

Steps to Understanding Sovereign AI Model Evaluation

• Benchmark Diversity: Government evaluators assess models across multiple dimensions, including long-context retrieval, mathematical reasoning, coding ability, and web research capability, rather than relying on a single metric.
• Architectural Trade-offs: Different design choices, such as LG's hybrid attention mechanism versus SK Telecom's sparse gated attention, create different performance profiles that suit different use cases and deployment scenarios.
• Training Methodology Impact: The choice to train in lower-precision formats like FP8 versus standard BF16, or to "upcycle" existing models versus training from scratch, affects both the final model's capabilities and the resources required to build it.
• Transparency in Limitations: Models that openly disclose their weaknesses, as SK Telecom did with A.X K2's web research performance, provide crucial information for enterprise and government buyers evaluating real-world deployment risks.

The August 8 through 11 evaluation will determine which two teams advance to the final phase of South Korea's sovereign AI competition. The release of these two models within 48 hours of each other signals that both LG and SK Telecom are confident in their technical approaches and ready to defend their positions in the government-backed race that aims to establish South Korea as a sovereign AI power independent of foreign technology dependencies.