On May 3, 2026, Seoul did something no government had done before with one of its own model labs: it bought in. The Korea National Growth Fund approved a 560 billion won (~$380.6 million) financing round for Upstage, the Seoul startup behind the open-weight Solar family of large language models. It was the fund's second direct AI bet, after a 640 billion won check into chip designer Rebellions in March, and the first time Korean public capital has taken an equity position in a domestic AI-model and software company rather than hardware (Korea Times).
The structure matters. Of the 560 billion won, 100 billion came from the Advanced Strategic Industries Fund and 30 billion from the Korea Development Bank, with the remaining 430 billion in private capital from SK Networks, Mirae Asset, Woori Venture Partners and others (Seoul Economic Daily). So roughly 130 billion won is a genuine state equity stake—public money on the cap table of a private model lab. Upstage is one of five teams—alongside Naver Cloud, SK Telecom, NC AI and LG AI Research—competing in the Ministry of Science and ICT's 'Independent AI Foundation Model' program (nicknamed Dokpamo), Korea's flagship sovereign-model push.
Betting on open weights, deliberately
What makes this more than industrial policy is what the state is funding. Upstage's flagship Solar-Open-100B is genuinely open-weight: a 102-billion-parameter mixture-of-experts model (12 billion active), pretrained on 19.7 trillion tokens, with weights downloadable from Hugging Face under the Upstage Solar License (Hugging Face model card). This is not a marketing gloss on a closed API. Developers can inspect, fine-tune, and self-host the model—the same openness that lets Korean SMEs, hospitals and ministries build on a domestic model without renting access from an American or Chinese lab.
Upstage is not an outlier. SK Telecom's A.X K1, a 519-billion-parameter sovereign model, shipped on Hugging Face under the permissive Apache 2.0 license (Computer Weekly). The pattern across the Dokpamo cohort is unmistakable: Korea is treating open release as a feature of sovereignty, not a risk to it. That instinct is sound. As former Hugging Face executive Tiezhen Wang argued this month, openly released models let downstream developers 'download, inspect, and deeply customize' rather than stay 'dependent on an American tech giant's ecosystem' (Rest of World). For a mid-sized economy, an open national model is leverage, not leakage.
The law that doesn't mention any of this
Here is the gap. Korea's Framework Act on AI—the AI Basic Act—took effect on January 22, 2026, and it is comprehensive on paper: risk-management duties for 'high-impact' AI in health, energy and public services; transparency and labeling for generative outputs; fines up to 30 million won; and a one-year enforcement grace period through 2026 (Future of Privacy Forum). The US Commerce Department's own summary confirms the Act reaches systems trained above 10²⁶ FLOPs (trade.gov).
What the Act does not do is say a single word about open-weight models or the pretraining data that builds them. There is no open-source carve-out, no clarity on who bears the labeling and risk-management duties once a model's weights are public and forked a thousand times, and no rule on the copyright or personal-data status of the trillions of tokens these models ingest. Seoul is now a shareholder in open models whose core regulatory questions its own statute leaves blank.
Steelman, then the fix
The case for moving slowly is real. Open weights are irreversible: once Solar-Open is downloaded, no recall is possible, and a downstream fine-tuner can strip safety guardrails. A regulator that wrote a heavy open-model regime in haste could saddle the very labs Korea is funding with duties they cannot discharge—because they no longer control the deployed copies. Caution is defensible.
But silence is not caution; it is uncertainty, and uncertainty is its own tax on innovation. The EU AI Act, for all its flaws, at least carves out genuinely open-source models from several obligations while keeping systemic-risk duties on the largest. California's AB 1856, after public backlash, is moving to explicitly exempt open-source operating systems from its age-assurance regime (EFF). Both jurisdictions concluded that open models need named treatment in the text. Korea has not yet done that work.
The proportionate path is narrow and pro-innovation. First, define open-weight models in the AI Basic Act's enforcement decree and place compliance duties on the deploying entity, not the upstream releaser who has relinquished control—so a state-backed lab can open-source without owning every downstream fork's liability. Second, give pretraining a safe harbor: a clear text-and-data-mining exception, paired with opt-outs and provenance logging, so Korean labs train on a settled legal basis instead of litigation risk. Third, use the one-year grace period to publish guidance, not to defer the question to 2027.
Seoul has made an admirable bet that open models are the right vehicle for AI sovereignty. The financing is the easy part. The law now has to catch up to the thing the money is buying.