Open-Source 700B Models Released Two Days Apart as Korea Nears Sovereign AI Elimination Round
South Korea’s two largest sovereign AI competitors released rival foundation models within 48 hours of each other this week, signaling 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 8–11.
SK Telecom published A.X K2 — a 688-billion-parameter open-weight model — on Hugging Face on July 29 per SK Telecom’s official announcement. LG AI Research followed with K-EXAONE 2.0, a 750-billion-parameter model, on July 31 per LG AI Research’s announcement. Both models use nearly identical architectural blueprints — Mixture of Experts (MoE) design, 262,144-token context windows, and Apache 2.0 licensing that allows any company to use or modify them commercially without restriction. The convergence of release timing and architecture type makes the differences between them — in training method, benchmark profile, and claimed capability — the defining question ahead of the government’s second-phase evaluation.
What MoE Architecture Actually Means for These Models
The headline parameter counts are real, but they describe a different thing than most readers 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 — “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 per NYU Shanghai RITS analysis.
This is a design choice, not a limitation. The core idea, which traces to a 1991 academic paper and was popularized at large scale by Google’s 2017 Sparsely-Gated Mixture-of-Experts work, 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 per IBM’s architecture explainer. The 256-expert pool gives the model a wider range of specialized representations to draw on than a 37B dense model would possess — but the compute burden per response is close to that of a 37B dense model. That tradeoff is why every major frontier lab — DeepSeek, Mistral, and now both Korean sovereign AI teams — has adopted MoE at scale.
A long-context add-on further distinguishes K-EXAONE 2.0. LG’s model 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 per LG’s model card. The result is a 262,144-token context window — roughly the equivalent of processing several full-length novels in a single pass — without proportional memory costs. A.X K2 reaches the same context length through a different mechanism: 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.
Where the Models Diverge: Training and Benchmarks
The most technically significant difference between the two releases is how they were built.
K-EXAONE 2.0 was produced through a method 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 per K-EXAONE 2.0’s model card. The 750-billion-parameter result is therefore not a model trained from scratch; it 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 the SwiGLU activation branches in each layer prevented the exploding activations that otherwise destabilize training in deeper networks.
A.X K2 took a 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 per A.X K2’s model card. 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 also 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.
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) per The Elec’s benchmark report. 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 per the Korea Herald. 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 per SK Telecom’s official announcement.
One number in A.X K2’s model card is worth close attention. 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 per NYU Shanghai RITS analysis. 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 and scripted support dialogues while struggling at open-ended web research is revealing something important about where its post-training effort concentrated — and that is precisely the kind of capability gap that government and enterprise buyers need to evaluate before deploying these models in agentic workflows.
How a Five-Team Race Became Four — and Why It Matters
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 per the Korea.net government press release. The government’s plan called for one team to be eliminated after the first evaluation in December 2025.
Instead, two were eliminated.
On January 15, 2026, MSIT announced that Naver Cloud and NC AI had failed to advance per the Korea Herald’s report on Phase 1 results. 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 requirement for full in-house training per Inven Global’s report on the elimination. The encoder weights were used “as-is — in a frozen state — without being updated,” MSIT said. Naver Cloud and NC AI chose not to apply to the supplementary call that followed.
A startup-led consortium was added to fill the gap. Motif Technologies, alongside KAIST and Trillion Labs, was selected in February 2026 after a supplementary evaluation that assessed its ability to independently design and implement key technologies per the Korea Herald’s February 20 report. That makes the current field four teams: LG AI Research, SK Telecom, Upstage, and Motif Technologies.
The same evaluation criteria that eliminated Naver Cloud — originality and full in-house training — now apply to judging the models released this week. LG’s upcycling approach uses LG’s own prior weights as the foundation rather than frozen external open-source weights, a distinction the government has so far accepted as compliant. Whether evaluators continue to view upcycling as sovereignty-compatible is one of the open questions going into the August panel.
LG Scored Highest in Phase 1; Both Giants Are Now in the Final Stretch
LG AI Research earned the highest score in the first-round evaluation: 90.2 out of 100, on a three-part assessment covering benchmark performance (40 points), expert panel review (35 points), and user feedback (25 points). Co-head Lim Woo-hyung framed K-EXAONE 2.0’s significance around the development process rather than the parameter count. “The significance is not simply that we built a model with a large number of parameters,” Lim said per the Korea Herald. “Korean researchers independently completed the entire development process. It means we have secured the capability to compete with global frontier models at the same scale.”
SK Telecom’s Kim Tae-yoon, head of foundation models, described A.X K2’s ambition in terms of deployment breadth. “We have advanced various functions so that it can be widely used from individuals’ daily lives to office environments and manufacturing sites,” Kim said per SK Telecom’s official announcement, and the company said it is already discussing expansion of A.X K2 into manufacturing, defense, and biotech applications.
Upstage — the third team in the current field — is a Seoul-based AI startup that topped the first-phase evaluation among observers for its comprehensive training transparency, publishing full training logs for independent verification per KoreaTechDesk’s Phase 1 analysis. Upstage’s Solar Open 2 model, whose technical report was released this week, uses a 1M-token context window and a tokenizer that outperforms both K-EXAONE and A.X K1 on Korean token efficiency, reaching 4.41 bytes per token on Korean text per the Solar Open 2 technical report. K-EXAONE scores 4.12 bytes per token on the same test, and A.X K1 scored 4.27 — meaning the startup’s model compresses Korean text more efficiently than either large consortium’s release, which translates directly to lower inference costs on Korean-language tasks.
The fourth team, Motif Technologies, has not released a large-scale public model ahead of the evaluation, operating with a smaller public footprint than the three larger players.
Apache 2.0: Why the Licensing Decision Matters Beyond Korea
Both LG and SK Telecom chose Apache 2.0 — the same permissive license used by Meta’s Llama series — for their new releases. The choice was deliberate.
Apache 2.0 allows any individual, startup, or corporation to download, modify, and deploy the model for commercial purposes without restriction, royalty, or requirement to release modifications. LG’s previous EXAONE model releases used more restrictive licensing terms; K-EXAONE 2.0 marks the first time the series is published under a fully permissive commercial license per The Elec’s licensing report. For developers and enterprises outside South Korea, this means both models can be integrated into commercial products immediately.
The commercial licensing decision also signals a strategic bet: LG and SK Telecom are not trying to build proprietary moats around these weights. The government competition requires models to remain accessible, and the broader ambition — for Korean AI to expand into global markets, as LG AI Research’s Lim put it — is better served by maximum adoption than by licensing revenue.
What Wins in August Will Help Power South Korea’s National AI Chatbot
The stakes of the December 2026 final evaluation are not abstract. South Korea’s Ministry of Science and ICT is simultaneously running a parallel program — “AI for All” — to provide every one of the country’s 51 million residents with free, unlimited access to a government-funded AI assistant per TechTimes’ coverage of Korea’s AI law and chatbot program. The program requires operators to run at least 50 percent of the system on certified domestic Korean AI models — which means the sovereign foundation model project’s winners will anchor the national service.
The government has invested heavily in the supporting infrastructure. South Korea’s cabinet approved approximately $5.7 billion for the sovereign AI stack through the National Growth Fund in May 2026 per UPI’s report on the investment, and MSIT’s 2026 AI budget is reported at $3.5 billion, a roughly 30 percent increase year-on-year per the U.S. Department of Commerce’s Korea Digital Economy guide.
The 200 citizen evaluators being recruited this week — applications close August 4 — will test the current four models on real-world usability and agentic tasks, with results factored into the August 8–11 second-phase evaluation per BigGo Finance’s report on citizen evaluator recruitment. The second-phase cut will reduce the field from four to three. By December, evaluators will select the final two model teams that will serve as the technical backbone for South Korea’s national AI infrastructure.
For a reader evaluating whether either model is worth deploying commercially today, the Apache 2.0 license removes every access barrier. Both models run on Hugging Face, both support ten or more languages, and both offer context windows long enough for enterprise document processing. K-EXAONE 2.0 is the stronger choice for Korean-language retrieval tasks across long documents. A.X K2 is the stronger choice for structured mathematical reasoning. Neither should be deployed in agentic workflows that require open-ended web research before users independently verify performance on BrowseComp-type tasks.
Frequently Asked Questions
What is the difference between K-EXAONE 2.0 and A.X K2?
Both are large Mixture of Experts AI models released this week by South Korean companies under Apache 2.0 open-source licenses. K-EXAONE 2.0, released by LG AI Research on July 31, has 750 billion total parameters (37 billion active per inference) and was built by expanding LG’s prior 236-billion-parameter model. A.X K2, released by SK Telecom on July 29, has 688 billion total parameters (33 billion active) and was trained from scratch in FP8 precision using SK Telecom’s proprietary Sparse Gated Attention architecture. K-EXAONE 2.0 leads on long-context retrieval benchmarks; A.X K2 leads on mathematical reasoning. A.X K2 scored last among its comparison set on BrowseComp (open-ended web research), which SK Telecom’s own model card disclosed per NYU Shanghai RITS analysis.
What does “Mixture of Experts” mean, and why do both models use it?
Mixture of Experts (MoE) is an AI architecture that divides a model into many specialized sub-networks — “experts” — and routes each token of input through only a small number of them during inference. For K-EXAONE 2.0, only 8 of 256 experts activate per token, meaning only about 37 billion of its 750 billion parameters are engaged for any given response. This allows a model to accumulate the knowledge breadth of a very large network while running at the inference cost of a much smaller dense model per IBM’s MoE explainer. Both LG and SK Telecom chose MoE because it is now the dominant architecture for cost-effective large-scale AI — DeepSeek, Mixtral, and other frontier models all use it.
What is South Korea’s sovereign AI competition, and who is still in it?
The Ministry of Science and ICT’s Sovereign AI Foundation Model project is a government-backed competition to build large AI foundation models using entirely domestic South Korean technology and data — no frozen weights from foreign open-source models allowed. Five teams were originally selected in August 2025 per the Korea.net government press release. Two were eliminated in January 2026 — Naver Cloud (for using frozen Alibaba Qwen encoder weights) and NC AI — and a startup consortium led by Motif Technologies was added in February 2026 per the Korea Herald’s February 20 report. The current field is four teams: LG AI Research, SK Telecom, Upstage, and Motif Technologies. The second-phase evaluation (August 8–11) will narrow the field to three; a final evaluation by year-end will select two winning teams whose models will help power South Korea’s national AI chatbot for its 51 million citizens per BigGo Finance’s citizen evaluator report.
Can I use K-EXAONE 2.0 or A.X K2 commercially right now?
Yes. Both models are available on Hugging Face and licensed under Apache 2.0, which allows commercial use, modification, and redistribution without restriction. This is a notable change for LG AI Research, whose previous EXAONE model releases used more restrictive terms; K-EXAONE 2.0 is the first in the K-EXAONE series to be fully open for commercial deployment per The Elec’s licensing report. Running either model requires significant hardware: A.X K2’s model card references a deployment setup using four high-end AI accelerator GPUs with roughly 656 gigabytes of aggregate GPU memory per A.X K2’s Hugging Face model card.
