South Korea releases a state-backed open model to cut reliance on foreign AI

Motif Technologies, a roughly 30-person South Korean AI subsidiary, released Motif 3, a large language model built entirely from scratch as part of “Dokpamo,” a South Korean government programme funding competing efforts to build a sovereign national AI foundation model. Entrants are required to train without incorporating any frozen weights from foreign systems — a rule intended to reduce the country’s reliance on models from the United States and China. Motif competed against LG AI Research’s EXAONE line, Upstage’s Solar models and SK Telecom’s own effort; on the third-party Artificial Analysis Intelligence Index, Motif 3 scored highest among the four in the programme’s second-phase evaluation, ahead of Upstage, SK Telecom and LG.

The model is a mixture-of-experts design with 314 billion total parameters and 13.2 billion active per token, trained on roughly 12.5 trillion tokens spanning web text, code, mathematics and multilingual data. Weights, technical report and a quantised variant were released under the MIT licence, giving outside developers commercial-use rights the earlier beta had withheld.

The programme’s stakes extend past a single model: Dokpamo’s final evaluation, due in December, will select the teams that become primary technical suppliers for a planned “AI for All” initiative giving South Korea’s roughly 51 million residents free access to a national AI assistant, with a requirement that at least half of the service run on certified domestic models. The contest is one of the clearer examples of a government treating frontier AI capability as a matter of national infrastructure to be built domestically rather than licensed from a handful of American or Chinese labs, a stance echoed elsewhere as governments weigh dependence on foreign AI providers against the cost of building competitive models in-house.