Whose Model Can We Call Our AI? — Sovereign AI in South Korea

How LLMs Are Trained — 1. Datasets · 2. Modeling · 3. Sovereign AI

“Our AI” carries more meanings than it first appears to. It might mean a model that speaks Korean well, one operated by a Korean company, or a Foundation Model trained from scratch on Korean data and GPUs. Once the phrase also asks whether the model can keep running if an external API shuts down, or whether a security issue can be inspected and fixed internally, sovereign AI becomes less about the model's nationality than about who has control.

Recent marketing copy often compresses all these layers into one phrase: “developed in-house.” A service fine-tuned from an open-weight model and a model pre-trained from random initialization can both be valuable engineering achievements, but they were not built in the same way. Using Open Source legally is one question; whether outsiders will regard the result as an independent Foundation Model is another. This article is not an argument that one route is purer. It begins from the need for precise language about what was actually built.

Where South Korea's Sovereign AI Stands in August 2026

The government's independent AI Foundation Model project began in August 2025 with five selected teams: NAVER Cloud, Upstage, SK Telecom, NC AI, and LG AI Research. In the first evaluation in January 2026, LG AI Research, Upstage, and SK Telecom advanced, while NAVER Cloud and NC AI were eliminated. Motif Technologies was subsequently added, so as of August 5, 2026, four teams—LG, Upstage, SKT, and Motif—have entered the second evaluation stage. As of the day this article was written, the government had not yet announced a confirmed result for that evaluation, so I will not presume a ranking or final selection. The project's launch and first-stage results can be found in the Ministry of Science and ICT's announcements of the selected teams and the first-stage evaluation results.

Even the models currently public reveal more than one approach. SKT A.X K1 explicitly describes itself as a large MoE model trained from scratch, and Upstage Solar Open 100B is another public model trained from the beginning. LG's K-EXAONE technical report describes an MoE model with 236B total and 23B active parameters, together with its long-context and multilingual abilities. Motif-3 Beta states that it was trained in-house and includes custom architecture components, though its current Beta model card's license conditions and distribution scope also need to be read. Meanwhile, SKT's A.X 4.0 clearly discloses that it is based on Qwen2.5 and further trained on Korean data. The former is a route toward independent pre-training; the latter uses proven open weights for rapid localization.

It would also be narrow to judge Korea's AI ecosystem solely by whether a model passes this government project. Models such as NAVER's HyperCLOVA X family continue to develop through real services and research, and each company has different disclosure policies and business goals. The government's criteria define an “independent Foundation Model” for a particular program; they are not an absolute scale that ranks every useful AI built in Korea. It is nevertheless natural to demand stricter evidence of technical independence when a project uses public funds and resources.

Borrowing an Architecture Is Different from Borrowing Weights

It is difficult to find a modern LLM untouched by the Transformer. Using ideas spread through papers and Open Source implementations—RMSNorm, RoPE, SwiGLU, MoE, or Multi-head Latent Attention—does not mean someone else trained the model. Science and engineering advance by reproducing and improving published ideas. As the Transformers documentation for A.X K2 explains that the model builds on the DeepSeek-V3.2 architecture with SKT's modifications, architecture lineage can be disclosed while the weights are trained separately.

If pretrained Base Weights are taken and subjected to Continued Pre-training or SFT, however, a substantial share of the resulting model's capabilities came from the original model's data and compute. This is not a criticism. It can be an entirely rational way to solve practical problems in Korean, law, manufacturing, or finance quickly, without repeating pre-training that costs hundreds of billions of won. But in that case, “a Korean-specialized model based on open weights” or “a model post-trained in-house” is technically more accurate than “a Foundation Model developed from scratch.”

Distillation makes the middle ground more complicated. Even if the Student was trained from random initialization, when most of its data consists of answers and reasoning produced by an overseas Teacher, its weight file is new but the lineage of its knowledge and behavior still depends heavily on the Teacher. Conversely, if training centers on human-produced local data and verifiable environments while the Teacher only assists with the quality of a small portion, that dependence is lower. Asking only “were the weights trained directly?” is therefore insufficient. We must also ask what the model learned from, and whether the process can be repeated without an external model.

This distinction became a real issue in the government's first evaluation. The Ministry of Science and ICT stated that using Open Source is itself an ordinary and permissible development practice, but external model weights must be capable of being initialized and trained independently, with the technologies and licenses used disclosed transparently. In NAVER's case, even though an external encoder could legally be used under the Apache License, the inclusion of its frozen external weights inside the model was cited as a reason it did not meet this program's independence criteria. The government's policy Q&A explains specifically that this was not a license-violation question, but a question of the minimum independence required by the program. It is a useful example of why legally permitted use and a convincing policy claim of an “independent model” are not necessarily the same.

“Developed In-house” Has Several Levels

If a team discloses its use of Open Source and complies with the license and attribution requirements, building a new product and training techniques on top of it can reasonably be called in-house development. The problem appears when the object of that phrase is omitted. Unless a company states whether it developed the data pipeline, post-training recipe, serving engine, or the entire Foundation Model, listeners naturally infer the broadest meaning.

The lower rows of this table are not technically less valuable than the upper ones. The value users actually receive may come from good search, secure data integration, evaluation, latency optimization, and incident response. Indeed, when a strong Base Model already exists, repeating enormous compute solely for the prestige of saying “we built it from scratch” requires justification in environmental and resource terms. Technical self-reliance is not the same as rebuilding every component domestically. What matters is knowing where the external dependencies are and whether there is an ability to replace them if they disappear.

Being Legally Permitted Is Different from Being Ethically Convincing

Permissive licenses such as Apache 2.0 and MIT broadly allow commercial use, modification, and distribution. Actual terms still need to be checked model by model, and rights in data, trademarks, and patents may remain separate. Even so, there is no reason to regard developing a derivative model in compliance with those terms as morally wrong. One purpose of an open ecosystem is to let others use and improve its work.

The ethical issue usually lies not in use but in representation. Trust erodes when a company implies that it built everything from scratch without naming the base model, erases the provenance of Teacher-generated data, or presents contributions from public research as proprietary inventions. This can happen without a legal violation. Models promoted with government support and national representation in particular need a higher standard of provenance than ordinary products. If a Model Card and Data Card disclose the Base Model, Teacher, main data categories, licenses, ranges of training tokens and compute, and independent evaluations, outsiders can compare claims with evidence.

Conversely, “from scratch” does not guarantee absolute purity. The data may include documents produced by other models, while the project may use public tokenizers and training frameworks, overseas GPUs and cloud infrastructure, and architectures proposed in papers. If a synthetic dataset's Teacher is undisclosed, random initialization alone cannot explain the lineage of the model's behavior. Transparency is ultimately not a declaration that nothing external was used. It is the work of drawing clear boundaries around what was used and where.

Sovereign AI Is Broader Than a Model File

I think sovereign AI is easier to evaluate when it is separated into at least five layers.

1. Data: Can the organization control collection rights and storage location, and perform deletion, correction, and audits?
2. Compute: Are there options for continuing training and inference through an external policy change, export control, or outage at a particular cloud provider?
3. Model: Does the organization possess and have the right to modify the weights, and does the license permit the intended distribution and commercial use?
4. Operations: Can serving, monitoring, safety patches, and rollbacks be handled independently in the domestic environment?
5. Governance: Can the training lineage and risks be audited and explained by an accountable party, and can the supplier be replaced?

Under these criteria, an open-weight model can sometimes offer greater operational sovereignty than a closed domestic model. A model may be trained locally, but if its weights and training information remain closed and only one provider can operate it, society as a whole has little ability to verify or replace it. Conversely, a model that began overseas may provide considerable service-level sovereignty if it can be self-hosted domestically, with control over data and deployment and the option to replace it with another model. Calling that an independent Foundation Model would still be a separate claim. The discussion becomes clearer only when independence of model origin and autonomy of operation are kept on different axes.

Then Can We Say “We Developed It Ourselves”?

My answer is close to “yes, but you have to finish the sentence by saying exactly what you developed.” If open weights are combined with Korean data and an in-house evaluation system to make an excellent model, that is unquestionably development. The team's own SFT and serving stack, domain data, and safety policies are real technical assets. But if the Base Weights came from elsewhere, the claim cannot be expanded to include independent pre-training. If the model depends heavily on synthetic data from a Teacher, that lineage should also be disclosed. Conversely, a model trained from the beginning cannot prove sovereignty with benchmark scores alone. The organization, data, and compute needed to operate it reliably, modify it, and train the next generation must remain in place.

The best way to persuade outsiders is not to use fewer qualifiers but to provide more evidence. The more a team discloses about weight initialization and whether there was a Base Model, architecture changes, categories and rights of data sources, use of Teachers, training logs and checkpoints, independent evaluations, and the inference environment, the more clearly its contribution can be seen without leaning on the word “in-house.” If trade secrets prevent full data disclosure, the public portion can still be expanded through a third-party audit, a reproducible small-scale recipe, or source-level proportions and rights categories.

For South Korea's sovereign AI effort to be convincing, it cannot stop at counting from-scratch models. It needs an ecosystem that lawfully secures Korean-language and domestic-industry data, operates GPUs, power, and data centers reliably, evaluates models, and takes responsibility for failures. Whether Open Source was used is not itself an ethical standard. More important is whether contributions are described honestly, external dependencies are understood and replaceable, and knowledge taken from the open ecosystem is returned to it.

If sovereignty is interpreted as the purity of never using anyone else's work, almost all modern software and research fails the test. A more practical standard, in my view, is this: without needing an outside party's permission, can we understand, operate, modify, audit, and, when necessary, replace the core system? When that ability is joined by transparency about the technology's lineage, we can move beyond a merely legal model and begin to speak of “our AI” in a way that society can also find convincing.

Previous: How Are LLMs Trained? — From Pre-training to Distillation and RL