Startups that received government support for graphics processing units (GPUs) are accelerating development of artificial intelligence (AI) models.

Officials and industry sources said on Wednesday that companies benefiting from the Ministry of Science and ICT's GPU support programme are developing diffusion-based transformer large language models (LLMs), Korean-language document parsing models and action AI agents.

AI startup Trillionlabs receives support for 80 Nvidia H200 GPUs through the National IT Industry Promotion Agency (NIPA) high-performance computing support programme. It developed a diffusion-based transformer LLM, Trida (Trida)-7B. The company emphasised that unlike existing autoregressive methods, the model processes entire sentences in parallel, independently implementing a next-generation architecture that global big tech firms such as Google Gemini are only beginning to try. Trillionlabs also released the model weights and inference code.

A Trillionlabs official said, "It is not easy for a startup to implement challenging technology with its own computing resources, but government support bridged that gap." The official added, "The NIPA team communicated frequently during development and fine-tuned the goals, and that was a key factor in success."

Generative AI company Posicube also receives GPUs through the same programme and is using them to develop a lightweight Korean-language vision-language model (VLM·7B). It is a document parsing model that analyses PDFs, PPTs, Word files and scanned images and automatically converts them into markdown, focusing on handling formats unique to domestic public and financial institutions that overseas models cannot process. After development is completed, it plans to embed the model in its solution, robi G, and supply it to the financial sector.

Commerce AI agent startup Inhance receives GPU support for nine months under the 2025 supplementary budget high-performance computing support programme and is focusing on upgrading its action AI model, ACT-2. ACT-2 is based on a large action model (LAM) that autonomously controls web interface functions such as clicking buttons, selecting menus and entering fields. An Inhance official said, "Based on GPU resources, we will strengthen our competitiveness in the global AI market."

NIPA, which operates the government GPU support programme, checks monthly with cloud service providers (CSPs) that directly provide resources to beneficiary companies. A NIPA official said, "Unlike support where the final outcome comes out, GPUs are largely used at intermediate stages of development, so beneficiary companies are sensitive about disclosing it externally." The official added, "We manage performance cases internally, but provide them only when an intention to disclose is confirmed through the media and other channels."

The Ministry of Science and ICT will first allocate 4,224 of the 10,000 GPUs secured through a supplementary budget to 159 projects. It will also launch an additional March application for the remaining 5,000 GPUs, including 4,000 for industry and 1,000 for academia. After allocation, it plans to monitor usage regularly and retrieve and reallocate GPUs if insufficient use or use outside the intended purpose is identified.