First-round assessment results will soon be released for 5 elite teams participating in a government-backed project to develop a homegrown foundation model.

The Ministry of Science and ICT will conduct the first evaluation of the project through Jan. 15 and is expected to eliminate 1 of the 5 consortia. It will also conduct evaluations every 6 months, cutting the field from 4 teams to 3 in June 2026 and selecting a final 2 teams in December 2026.

Upstage, SK Telecom, NC AI and the LG AI Research team, which were selected among the 5 elite teams, aim to survive and make the final 2. For now, their first priority is to avoid being the first to be eliminated.

Perhaps because of that, tensions among the companies ahead of the first evaluation are also running high.

Recently, Upstage, Naver Cloud and SK Telecom, among the 5 elite teams, were swept up in controversy over “from scratch” development, meaning independently building AI models from the beginning. After the 5 teams held public technical presentations at a Ministry of Science and ICT-hosted event on Dec. 30, smaller and larger disputes over from-scratch development emerged.

The disputes have flared because from-scratch development is a first-round evaluation criterion in the foundation model project. The ministry has stressed that developing derivative AI models through steps such as fine-tuning overseas AI models does not count as developing a domestic AI foundation model.

In Upstage’s case, the controversy arose after Seok-hyun Ko, chief executive of AI startup Psionic AI, raised suspicions on the developer platform GitHub that “Solar Open 100B” is a derivative model based on Chinese Zhipu AI’s “GLM-4.5-Air” model.

Soon after the allegations, Upstage moved to put out the controversy through public verification. Ko also posted an apology on social media, bringing the dispute to a close for now.

SK Telecom faced criticism that A.X K1, which it is developing as a foundation model, is partly similar to a DeepSeek model. The company said the issue was related to execution code and was unrelated to building a model from scratch.

Naver Cloud introduced as domestic foundation models a foundation model applying a native omnimoidal structure, “Native OmniModel (HyperCLOVA X SEED 8B Omni),” and a “high-performance reasoning model (HyperCLOVA X SEED 32B Think)” that adds visual, audio and tool-use capabilities to existing reasoning AI.

The company said the Native OmniModel is the main model for the foundation model project. But controversy arose over the HyperCLOVA X SEED 32B Think unveiled alongside it, related to its use of fine-tuned weights from the Qwen 2.5 vision encoder.

Naver Cloud explained that it has developed the core engine of the foundation model 100% with its own technology from the From Scratch stage, securing competitiveness in deeply understanding the Korean language and the complex context of Korean society. It said it strategically adopted a proven external encoder for this model, considering compatibility with the global technology ecosystem and efficient optimisation of the overall system.

It also stressed that the vision encoder serves as an “optic nerve” that converts visual information into signals the model can understand, and that Naver has sufficient proprietary vision technology capabilities, including VUClip.

Some see the recent controversy as the result of overheated jockeying among companies ahead of the first evaluation, going beyond natural questions raised at the community level. Others point to a lack of detailed government-level criteria for From Scratch development, allowing disputes to spread excessively.

Concerns are also emerging that lingering aftereffects will remain even after the first evaluation. An industry official said there is substantial room for interpretation in the criteria, adding that practices accepted globally in AI development are becoming a source of controversy and that there is a need to strengthen expertise in evaluations.