Sovereign AI means independently developing and operating the AI that best understands a country or region’s own language, culture, and values. As AI rapidly reshapes industries and daily life, countries around the world are competing fiercely to take the lead. Through all of this, NAVER Cloud has consistently emphasized the importance of sovereign AI and worked to strengthen Korea’s domestic AI ecosystem.
Now we’re taking another meaningful step forward: joining the Nemotron Coalition led by NVIDIA. In this post, we’ll look at what the Nemotron Coalition is and what joining it means for us.
What is the Nemotron Coalition?
The Nemotron Coalition is a global collaboration in which open-model builders and AI developers share research, expertise, data, and compute to advance open frontier models. NVIDIA announced its launch at GTC 2026, with inaugural members drawn from leaders across the field—Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam, and Thinking Machines Lab. The goal is to bring the world’s foremost AI labs and infrastructure providers together, pooling capabilities that were once scattered to accelerate the development of open frontier models. This approach lowers the barrier to cutting-edge AI innovation and helps developers and companies everywhere build and refine models for their own industries, regions, and use cases.
Each member has a clear role. Mistral AI and NVIDIA jointly develop the base model, and coalition members contribute their own data, evaluations, and domain expertise to support post-training and ongoing development. The resulting models are trained on NVIDIA DGX Cloud, shared back into the open ecosystem, and will form the foundation of the upcoming Nemotron 4 family.
As member companies advance these models with their own expertise, NAVER Cloud contributes its Korean and multilingual data, along with hands-on experience training and operating hyperscale models.
Why build frontier models together?
Today’s AI no longer just answers a question and stops. It plans its own next steps, calls the tools it needs, checks the results, and tries again—rapidly evolving into what’s known as a “long-running agent.” Carrying a complex, multi-step task through to completion takes a model that can explore more possibilities in less time and produce more accurate results, faster.
The challenge is that frontier models like these are becoming harder for any single company to build alone. A strong model demands vast amounts of data, a stable training environment, a rigorous evaluation framework, and real-world domain knowledge all at once. That’s why collaboration that pools the world’s capabilities matters more than ever—and it’s exactly why NAVER Cloud joined the Nemotron Coalition.
Introducing Nemotron 3 Ultra
The model at the heart of this partnership was unveiled at GTC Taipei 2026, where CEO Jensen Huang presented NVIDIA’s new open model, Nemotron 3 Ultra, on the keynote stage. Nemotron 3 Ultra is built for exactly the kind of long-running agents described above. With a Mixture-of-Experts (MoE) architecture of 550 billion parameters, it handles the heaviest decisions in an agentic workflow—making architectural choices over long coding sessions, synthesizing hundreds of research sources, and verifying thousands of interlocking constraints.
It’s worth pausing on why an MoE architecture matters here. The idea is to keep the total parameter count large while activating only a subset of experts for each token. Of the 550 billion total parameters, only about 55 billion are active at any given moment. That lets the model balance the reasoning quality of a large model against an inference cost you can realistically afford—and the more a system relies on running long agent loops over and over, the more directly that efficiency translates into practical, real-world value.
Advancing HyperCLOVA X
This brings us to the core of the partnership: advancing HyperCLOVA X. Drawing on the technology behind NVIDIA’s open LLM, Nemotron 3 Ultra, we’re developing the next generation of HyperCLOVA X. Beyond that, the two companies plan to jointly research the optimization and foundational technologies of hyperscale language models.
What makes this partnership compelling is how the two companies’ strengths complement each other. NVIDIA brings world-class model-development expertise and a state-of-the-art hardware and software stack, while NAVER Cloud brings Korean and multilingual data along with hands-on experience training hyperscale models and running them as live services. It’s as if one side shows up with the world’s best materials and tools, and the other with deep knowledge of the Korean landscape and the know-how to operate within it.
Our aim through this partnership is clear: to deliver, faster than anyone, a top-tier AI model built for the Korean language and Korean culture. On the solid foundation of a global open model, we’re adding Korean-language, coding, math, and reasoning capabilities, and optimizing them for our own GPU environment and service requirements.
More than that, this partnership goes well beyond simply taking a finished model and putting it to use. Together with NVIDIA, we’re designing and advancing the underlying technology itself. In practice, as we train the models, we share technical suggestions for improving the architecture directly with NVIDIA. In other words, we’re not just a user—we’re a co-developer, contributing directly to the model’s progress and building open frontier models together.
Sovereign AI: Why this partnership matters to us
There’s a clear reason this partnership means so much to us: it aligns precisely with the value of sovereign AI.
Nemotron models are released with open weights, datasets, and recipes. That means you’re not handed a model as an inscrutable black box—you can look inside it, keep control, and tailor it to your own domain. Above all, you can deploy it right where your applications and data already live. The ability to run a frontier model without sending data outside is a decisive advantage in fields with demanding requirements around regulatory compliance, data sovereignty, and data localization. This is exactly where open frontier models meet the promise of sovereign AI: building and running AI on your own data and infrastructure.
Closing thoughts
Joining the Nemotron Coalition and advancing HyperCLOVA X are really just the starting point of a long effort to build open frontier models together. We want to contribute to the open ecosystem while creating a virtuous cycle that brings those gains back to our own services and users. As we move forward step by step on the path to sovereign AI, shoulder to shoulder with the world’s best partners, we hope you’ll keep following NAVER Cloud’s progress.