South Korea's government-backed AI competition is revealing an unexpected truth: building world-class artificial intelligence doesn't require thousands of engineers or unlimited budgets. As the country's Dokpamo sovereign AI program enters its final evaluation phase, smaller teams are producing results that rival models from the country's largest tech conglomerates, suggesting that strategic architecture and focused training can outweigh raw scale.

How Are Smaller AI Teams Competing With Tech Giants?

The competition, which closes its citizen evaluator applications this week before an August 8-11 elimination round, has drawn international attention primarily to two heavyweight contenders: LG AI Research's K-EXAONE 2.0 with 750 billion parameters and SK Telecom's A.X K2 with 688 billion parameters. But the real story lies with the two other finalists, whose performance on independent benchmarks suggests that size alone doesn't determine capability.

Motif Technologies, a subsidiary of chip company Moreh with roughly 30 employees, released Motif-3 Beta on July 14. On the Artificial Analysis Intelligence Index, which aggregates results across nine evaluations covering reasoning, knowledge, coding, and mathematics, Motif-3 scored 44, placing it 34th among 577 tracked models globally. That puts the model in the global top 7% despite having far fewer engineers than competitors.

What makes this achievement notable is how Motif Technologies used its government-allocated compute resources. The company received approximately 768 NVIDIA B200 graphics processing units (GPUs), the same amount provided to larger competitors. Rather than simply scaling up a standard architecture, Motif-3 uses a Mixture of Experts design with only about 13 billion parameters active per inference call, among the lowest active-parameter counts in the field.

The model's architecture relies on two proprietary mechanisms: GDLA (Gated Differential Latent Attention), which modifies how the model processes information at the latent representation stage, and a "polynorm" per-expert activation function that replaces standard normalization inside each expert sub-network. Motif Technologies describes the overall architecture as "wholly in-house," not adapted from existing open-source models.

What Design Choices Allow Smaller Models to Punch Above Their Weight?

Upstage, a Seoul-based AI startup, took a different approach with Solar Open 2, released on July 22. Despite activating only 15 billion of its 250 billion total parameters per inference call, Solar Open 2 scored 86.8 on Ko-GDPval, a Korean-language benchmark that tests a model's ability to produce complete office documents from task descriptions. This score essentially matches DeepSeek-V4-Pro, a 1.6-trillion-parameter model more than six times its active size.

Solar Open 2's efficiency comes from a fundamental architectural rethinking. Most large language models use Rotary Position Embedding (RoPE) to tell the model where each token sits in a sequence, but this approach requires workarounds when extending context length during inference. Upstage removed positional encoding entirely, a design called NoPE (No Positional Encoding), and instead relies on the model's linear-attention layers to encode token order intrinsically in their recurrent hidden state.

The result is a 1-million-token context window, the largest among the four Dokpamo finalists, that runs on two NVIDIA H200 GPUs with quantization applied. On APEX-Agents, the leading benchmark for agentic task performance, Solar Open 2 scored 16.6, the highest in its comparison class, ahead of MiMo-V2.5 at 13.4 and DeepSeek-V4-Flash at 13.2.

Upstage developed an OfficeVerse simulation environment, purpose-built synthetic scenarios covering document writing, spreadsheet analysis, presentation creation, and coding, to teach the model to complete real-world tasks rather than simply predict the next word. A Multi-teacher On-Policy Distillation approach consolidated specialized skills from multiple task-specific models into the single Solar Open 2 architecture.

Steps to Understanding Sovereign AI Competition Criteria

• Round 1 Priorities: The first evaluation phase emphasized benchmark performance (40 points), expert panel review (35 points), and user feedback (25 points), with significant emphasis on model originality. This criteria eliminated Naver Cloud for submitting a model that incorporated frozen encoder weights from Alibaba's Qwen model in January 2026.
• Round 2 Shift: The second round shifted the balance toward real-world applicability and agentic capability, with all four teams working to demonstrate deployment use cases in manufacturing, finance, legal services, healthcare, and public administration.
• Deployment Focus: SK Telecom developed quantization-optimized models for South Korea's Ministry of National Defense and is demonstrating manufacturing AI agents with steel producer KG Steel, while LG AI Research highlighted a 30% improvement in agentic coding performance in K-EXAONE 2.0 versus its predecessor.

The shift in evaluation criteria reflects a broader recognition that sovereign AI isn't just about achieving the highest benchmark scores. It's about building systems that can actually work in government agencies, critical infrastructure, and enterprise environments where reliability and local control matter.

Why Does Sovereign AI Security Matter Beyond Competition Rankings?

While South Korea's competition focuses on capability and real-world deployment, the broader sovereign AI movement globally is increasingly intersecting with cybersecurity concerns. Nvidia has launched the Open Secure AI Alliance, an industry coalition bringing together technology and cybersecurity companies to develop and share open tools, models, and techniques for AI security.

The initiative builds on the Linux Foundation's Akrites initiative and work by the Open Source Security Foundation (OSSF), focusing on using open technologies to identify, remediate, and disclose vulnerabilities. Nvidia argues that cyber defenders need access to AI systems they can inspect, modify, and deploy themselves rather than relying exclusively on proprietary platforms.

This argument carries particular weight in the Middle East, where countries including the United Arab Emirates and Saudi Arabia are investing heavily in sovereign AI capabilities while simultaneously strengthening national cybersecurity frameworks. Abu Dhabi-based G42 is among the companies joining the alliance, alongside Microsoft, IBM, Cisco, CrowdStrike, Cloudflare, Dell Technologies, Hugging Face, Palantir, and the Linux Foundation.

"It could shift AI security from relying on proprietary solutions to adopting open, transparent and collaborative security tools," said Irene Corpuz, founding partner of Women in Cyber Security Middle East. "This would enable organisations to better audit, customise and validate AI systems while becoming more resilient through shared standards and community-driven innovation."

Irene Corpuz, Founding Partner of Women in Cyber Security Middle East

For Gulf organizations balancing rapid AI adoption with national requirements around cybersecurity and sovereignty, the ability to deploy security models locally could become increasingly important for government agencies, critical infrastructure operators, and companies handling sensitive data.

Meanwhile, China is taking a different approach to protecting its sovereign AI capabilities. The Chinese government announced revised regulations for the protection of integrated-circuit designs, intended to safeguard the intellectual property of domestic AI chip breakthroughs. Premier Li Qiang signed a revision to the Regulations on the Protection of Layout-Designs of Integrated Circuits law, originally enacted in 2001, to expand its scope to reflect today's chip market.

The revisions include widening the definition of integrated circuits protected under the regulation to include those with "photonic, quantum, or other functions." Additional changes outline specific rates of remuneration owed to individuals whose chip designs have been commercialized, as well as a "punitive damages system" for design infringement.

The timing reflects China's broader push toward technological self-sufficiency. Chinese chip giants including Huawei are seeing significant growth, with the Financial Times reporting that Huawei is set to reap a 60% increase in chip sales this year. ChangXin Memory Technologies (CXMT), a DRAM memory manufacturer, saw a 470% surge in shares last week as it made its Shanghai Stock Exchange debut. In July, the AI firm DeepSeek announced plans to develop its own AI chips to reduce dependence on Nvidia and Huawei chips.

These developments across South Korea, the Middle East, and China illustrate a fundamental shift in how nations approach AI development. Rather than relying on a small number of global technology providers, countries are investing in domestic infrastructure, building homegrown models, and protecting their intellectual property. The success of smaller teams in South Korea's competition suggests that with the right strategic focus and access to computing resources, nations can build competitive AI systems without depending on foreign technology giants.