KT announced on the 9th that its self-developed artificial intelligence (AI) “Mi:dm 2.0” ranked first in the domestic performance evaluation platform “Horangi Leaderboard” for Korean large language models (LLM).
The “Mi:dm 2.0,” which KT unveiled on the 4th, consists of two types: the “Mi:dm 2.0 Base” with 11.5 billion parameters and the “Mi:dm 2.0 Mini” with 2.3 billion parameters. Among these, the base model recorded the best overall performance among domestically developed models with fewer than 15 billion parameters in the Korean LLM evaluation metric “Horangi Leaderboard3.” This model achieved a general performance score of 0.7197, with a universal performance score of 0.7004 and an alignment performance score of 0.739. It ranks third among similar models globally.
“Horangi Leaderboard” is a Korean-language-specific LLM evaluation benchmark organized by the global machine learning operations (MLOps) company Weights and Biases (W&B). It comprehensively evaluates the practical capabilities of Korean LLMs, including general language comprehension and response safety in actual Korean environments, such as alignment. Unlike existing benchmarks that assess accuracy in answering questions with fragmented knowledge, it is characterized by evaluating based on real-life scenarios, including writing, reasoning, and information extraction.
It is considered an important standard for gauging the model’s effectiveness, as it can also assess the unique contexts, expressions, and social contexts of the Korean language. The evaluation results are transparently published through a platform operated by W&B, allowing anyone to verify them directly.
KT stated regarding this evaluation, “Mi:dm 2.0 objectively demonstrated that it possesses practical competitiveness as a Korean AI model based on domestic technology.” They noted, “Mi:dm 2.0 was developed to reflect Korea’s spirit and lifestyle, knowledge, and social context, making it the most suitable AI for Korea.”
“Mi:dm 2.0 was developed using a ‘from scratch’ method, where KT directly carries out the entire process, from architecture design to data construction and learning, rather than simply fine-tuning based on foreign models.
KT plans to provide customized business-to-business (B2B) AI solutions based on Mi:dm 2.0 across various industries and will promote empirical applications in public, financial, educational, and legal sectors. They are also preparing to release an upgraded version of Mi:dm 2.0 Pro, which enhances performance. Additionally, they plan to strengthen the AI ecosystem through collaborations with domestic AI semiconductor company Rebellions, along with expanding the reasoning and multimodal capabilities of Mi:dm 2.0.
Shin Dong-hoon, head of KT’s Generative AI Lab (Chief AI Officer), stated, “We will actively promote the market diffusion of ‘Korean AI’ by applying Mi:dm 2.0 to various domestic industries,” and emphasized, “We will position ourselves as a key player driving the practicalization and technological advancement of Korean AI.”