KT to release homegrown AI model Mi:dm 2.0 as open source

KT officials test the Mi:dm 2.0 large-language model at the company's R&D center in Seocho District, Seoul, in this handout photo, released Thursday. Courtesy of KT

Mobile carrier KT said Thursday it will release its self-developed artificial intelligence (AI) model Mi:dm 2.0 on developer platform Hugging Face as open source, allowing companies and individuals to use it commercially without restrictions. The name Mi:dm is derived from the Korean word "mideum," meaning belief.

Mi:dm 2.0 is an upgraded version of KT’s self-developed large-language model (LLM), which reflects KT’s philosophy of Korean AI. The model is built on Korea’s values, ways of thinking and knowledge, incorporating intangible elements such as social context and cultural characteristics unique to the Korean language.

KT first unveiled standard and premium versions of Mi:dm 1.0 in 2023, and has been adopting it for its AI customer center, Genie TV, AI phone assist and call centers.

Mi:dm 2.0 comes in two versions — a 11.5 billion-parameter Base version and a 2.3 billion-parameter Mini version. Both versions support Korean and English.

Mi:dm 2.0 Base is a general-purpose model optimized for Korean-specific knowledge and document-based question answering. Mi:dm 2.0 Mini is a smaller model trained by knowledge distilled from the Base model.

By releasing them as open source, KT will be the first Korean company to open-source a Korean-language LLM with more than 1.1 billion parameters for unrestricted commercial use.

KT said the Mi:dm model has demonstrated superior performance in comprehension and generation in fields related to Korean language, culture and society compared to leading domestic and global models. In the Ko-Sovereign benchmark, an AI evaluation metric co-developed by KT and Korea University to measure AI proficiency in Korean-language tasks, Mi:dm outperformed similar-sized domestic models and top-tier global open-source models.

Mi:dm also outperformed major domestic and global open-source models in KMMLU, a benchmark for evaluating Korea-related knowledge, and HAERAE, a Korean-language model assessment metric.

KT also worked closely with domestic AI processor maker Rebellions during the development of Mi:dm to optimize its performance on domestically produced AI chips.

By open-sourcing Mi:dm 2.0, KT said it will promote the philosophy of Korean AI in the domestic ecosystem. In collaboration with Microsoft, the company also plans to release a model that trains GPT-4 with Korean ways of thinking.

“Mi:dm 2.0 is a sophisticated AI model designed to show general generative capabilities based on a deep understanding of Korean language and culture,” KT Gen AI Lab head Shin Dong-hoon said.

“KT will offer a new alternative for high-performance Korean AI models and Mi:dm 2.0 will serve as a key stepping stone for KT’s global competitiveness.”