Proprietary development of the ‘Ivy’ platform based on ultra-large-scale generative AI
Clevi Co., Ltd. (Clevi, CEO Lee Hwan-ho) announced that it would begin the open beta service for ‘Ivy’, an ultra-large-scale generative AI developed in-house using LLM (large language model) technology.
Domestic start-up Clevi has officially launched its own ultra-large-scale generative AI model, ‘Ivy’, using LLM (large language model) technology. Until now, South Korea had not achieved notable results in the AI market amid the dominance of OpenAI in the United States and DeepSeek in China. However, by officially launching ‘Ivy’, Clevi has demonstrated that South Korea can also develop AI using its own technological capabilities.
Ivy provides powerful features including △real-time learning and adaptation △multi-data processing △a code interpreter △expert model creation and △an agent support system, enabling it to be used across a wide range of fields, from individual users to businesses.
Over a period of more than two years, Clevi CEO Lee Hwan-ho developed the proprietary model from pre-training onwards through an efficiently designed training pipeline. According to the company, the model cost approximately KRW 1.5 billion to develop, while operating costs average around KRW 10 million per month.
CEO Lee emphasises that “‘Ivy’ has high efficiency and high performance, enabling it to solve the problem that has been identified as a shortcoming of generative AI: ‘PoCs are easy, but completion is not’.” This means that businesses can reduce labour costs and improve productivity by having AI perform repetitive and inefficient tasks on their behalf.
Clevi stated, “To minimise AI development costs, we introduced innovative technology that efficiently manages the development process by applying automated container-based training node deployment technology,” adding, “We significantly reduced costs through the efficient design of the training pipeline and optimisation of the communication network between nodes.”
According to Clevi, a feature of the ‘Ivy’ model is the development of Mixture of models for efficient model training and operation. It was developed by creating small models, then distilling and combining those models into the model, and supports not only English and Korean but also a wide range of languages worldwide through multilingual training and embedding optimisation.
It is explained that, amid the fierce competition in AI technology, with OpenAI in the United States and DeepSeek in China competing to take the lead, Korea’s homegrown large-scale generative AI ‘Ivy’ has attracted industry attention by recording higher scores than the two existing models in various benchmark tests (△AIME 2024: 81% △MATH-500: 98.6 △SWE-BENCH: 51.1 △MMLU: 91.2).
CLEVI has made the Ivy app platform available for registration with minimal personal information, amid growing concerns about personal information leaks and security as the use of AI assistants expands. It has also been recognised for resolving security issues by supporting on-premises installation for public institutions and companies to suit their user environments.
A CLEVI representative said, “Our company will not stop at the launch of Ivy; we plan to release an agent that will directly intervene, respond and participate in conversations between individual and group users in the future. We have also completed research and development on ARI (Adaptive Reasoning Intelligence), an AI-based active workflow agent, as well as methods for precise object recognition and real-time control in video content production and medical image analysis, and are preparing to make these services available,”
Meanwhile, ‘Ivy’ has been registered on the Google Play Store, and users can also download and use Ivy from the company’s official website and the Ivy download page.
Choi Yong-seok, Donga.com reporter [email protected]
Source: Donga Ilbo (reporter Choi Yong-seok) | https://www.donga.com/news/It/article/all/20250325/131275337/1