Open Sourcing a Newly Developed AI Model
Upgraded Korean-Specialized LLM
Responding to Government's 'Sovereign AI' Policy with an Independent Model
KT announced on the 3rd that it plans to release the open source of its language model "MidEum 2.0," which embodies the philosophy of "Korean-style AI," through the AI developer platform Hugging Face. The photo shows KT researchers from the Technology Innovation Division testing MidEum 2.0 at the KT Umyeon Research Center in Seocho-gu. Photo by KT
KT is making a bold move by releasing its self-developed Korean-specialized large language model (LLM), "MidEum 2.0," as open source. This aligns with the domestic trend toward "sovereign AI." KT, which had previously indicated a strategy focused on collaboration with big tech companies, appears to be shifting its approach. The company is now increasing the utilization of its independently developed model in line with the government's recent push for AI independence.
On July 3, KT announced that its generative AI research division, Gen AI Lab, will release the newly developed MidEum 2.0 on the global open source platform Hugging Face. The model will be available for anyone to use commercially without restrictions and comes in two versions: "MidEum 2.0 Base" with 1.15 billion parameters and "MidEum 2.0 Mini" with 230 million parameters. Both versions support Korean and English.
MidEum 2.0 is the next-generation model following the 1.0 version introduced by KT in 2023. It features significant advancements in parameter size, training data, and Korean language processing capabilities. According to KT, the new model offers greatly improved versatility and performance compared to its predecessor.
KT emphasized that this release is more than just a technical sharing initiative. The company aims to spread a model that deeply understands the Korean language and the context of Korean society, based on the philosophy of "Korean-style AI." In fact, the model demonstrated superior performance compared to domestic and international open source models in the "Ko-Sovereign" benchmark, a Korean-specialized evaluation metric jointly developed with Korea University.
This move signals a shift in KT's AI policy direction. Recently, as the government adopted sovereign AI as a core policy and emphasized strengthening national AI competitiveness, KT has also begun to focus more on the external dissemination and utilization of its proprietary AI models. Previously, KT prioritized partnerships with big tech companies such as Microsoft (MS). Internally, there was a prevailing view that leveraging big tech's technological capabilities to enhance added value was a more realistic strategy than independent development. However, KT had taken a cautious approach, temporarily releasing its first version of the proprietary LLM, MidEum, and the "MidEum 7B" (700 million parameter model) on Hugging Face before making them private again after a few months. This led to industry criticism that KT's efforts were "stagnant."
The newly released MidEum 2.0 is characterized by its high precision and language comprehension, having been trained on a diverse range of Korean-specialized data, including literature, law, patents, and dictionaries. KT developed its own tokenizer optimized for the structure of the Korean language and addressed copyright issues to enhance ethical standards and transparency. Additionally, KT worked closely with the startup Rebellions during development to optimize the model for operation on domestic AI semiconductors. In collaboration with Friendly AI, KT is also temporarily providing an environment where users can try the model for free on Hugging Face without a separate installation process.
Shin Donghoon, head of KT Gen AI Lab (CAIO), stated, "MidEum 2.0 is an advanced model that not only possesses general generative capabilities but also deeply understands the Korean language and culture. It will provide a practical alternative for domestic users and serve as a foundation for securing global competitiveness."
KT plans to sequentially introduce additional models, such as a GPT-4-based model reflecting Korean perspectives, in collaboration with MS, starting with this release. The company aims to focus on building an AI ecosystem led by private enterprises while also fine-tuning its direction in coordination with the government.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
