To be released as open source on HuggingFace on July 4
"Deep understanding of Korean language and culture"... Designed for B2B specialization
Setting a standard for 'Sovereign AI'... "Collaboration with MS to continue"
KT announced that it will release its self-developed Korean-specialized large language model (LLM) "Mi:um 2.0" as open source starting July 4, and that it intends to actively participate in the government's independent artificial intelligence (AI) foundation model development project. This marks a shift from KT's previous focus on collaboration with global big tech companies such as Microsoft (MS), as the company is now placing greater emphasis on developing its own models.
On July 3, KT held an online briefing and stated, "We are releasing Mi:um 2.0, which we have developed over the past year with our own technology, on HuggingFace so that anyone can use it," adding, "We are also preparing to participate in the government's independent AI foundation model project."
On this day, Shin Donghoon, head of Gen AI Lab, explained in his presentation, "KT has never stopped developing its own technology," and added, "Mi:um 2.0 was developed with a tokenizer specifically designed for Korean, focusing on features that are highly useful in B2B environments such as document understanding, report writing, and document-based Q&A."
The model lineup is expected to expand beyond the Mini and Base versions to include high-performance Pro models, inference models, and multimodal models. He stated, "We made efforts to reflect Korean nuances, emotional expressions, historical perspectives, and etiquette," emphasizing, "This is not just a generative AI, but a model that embodies Korean values and practicality."
Shin stated, "Because KT handled everything from model design to training without any external assistance, we are confident that this is a representative model of sovereign AI." He also emphasized the "RAI (Responsible AI)" principle, a technology for correcting or removing information learned by the model that does not correspond to facts. KT explained that RAI governance was applied throughout the entire development process of Mi:um 2.0.
The performance of the model is supported by data. KT collected its own data and also utilized Korean-language data secured through the K-Data Alliance, classifying and refining it into over 200 categories for training purposes. When asked about copyright issues, KT responded, "All data was secured through purchase or partnerships, and any data in a gray area with ambiguous licensing was strictly excluded from training."
In response to the question, "Why develop the Mi:um model separately while also building MS and GPT-based models?" KT explained, "Not every task requires a GPT-level model, so we provide both models complementarily depending on the purpose and usage environment." While GPT-based models are described as "heavy and powerful models" suitable for tasks that require high performance, such as complex reasoning or creative work, Mi:um 2.0 is characterized as a relatively lightweight, practical model based on domestic technology, designed for a deep understanding of the Korean language and culture and offering high cost-effectiveness.
Regarding the possibility of repeating the precedent of making "Mi:um 7B," which was uploaded to HuggingFace in 2023, private again, KT stated, "This release is a commitment to genuinely contribute to the AI community as open source," and emphasized, "This is not a time-limited release; we plan to continue to advance the model over time."
Meanwhile, the specific participating companies and the scale of data secured through the K-Data Alliance are scheduled to be announced in mid-July. KT also plans to provide separate briefings on business models, revenue plans, and mid- to long-term investment strategies.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
