KT to launch large language model for enterprise use

Industry

Published: 31 Oct. 2023, 17:11

KT to launch large language model for enterprise use

Second from left, KT's Choi Joon-ki, head of AI & big data business unit, Bae Soon-min, head of AI2XL research center, a local education startup Mathpresso CEO Jake Lee another AI startup Upstage CEO Kim Sung-hoon participate at an online press event on Tuesday to introduce KT's large language model Mi:dm. [KT]

KT will release its own large language model called “Mi:dm” this year. The telecommunications provider hopes to strengthen its enterprise generative AI business in five areas: manufacturing, finance, education, the public sector, and the global market.

As of Tuesday, KT is accepting applicants for its first wave of Mi:dm users on its webpage. It aims to publicly release Mi:dm's foundation model through a dedicated portal site by the end of 2023.

Mi:dm is categorized under four different parameter scales, ranging from a minimum of 7 billion parameters to a maximum of 25 billion. The 7 billion-parameter model, dubbed the basic model, ranked No. 1 on Hugging Face’s Open LLM Leaderboard, an evaluation platform which grades large language models on standards such as reasoning, common sense interference, contextual understanding and factual accuracy.

“Mi:dm’s foundation model will be publicly released for free on KT’s portal site ‘KT Mi:dm studio,’ free to be utilized for all clients,” said KT’s Choi Joon-ki, head of AI and big data business in an online press briefing Tuesday. “Clients who wants to deploy the large AI but do not have the capacity to create multi-billion-parameter models will be able to fine-tune the Mi:dm model to fully adapt the software into their corporate system.”

KT said that it's been able to reduce Mi:dm's problems with hallucination — in which a generative AI creates misleading or inaccurate content — by 70 percent.

Mi:dm incorporates three software programs, the company added, including “Document AI,” which distills complicated texts to a level the model can easily comprehend; “Search AI,” which uses deep learning technology to find current data related to queries; and “Factguard AI,” which uses reinforcement learning skills to help the LLM form its answer.

KT will also offer clients its subscription-based “hyperscale AI computing” service, hosted by subsidiary KT Cloud, which leverages a GPU farm to boost a device's computing capabilities. The company says that its service can reduce the cost of training its model by approximately 27 percent — or by 50 percent if a client deploys a neural processing unit developed by Rebellions, a local AI chip designer that KT has previously invested in. Clients will pay only for the specific cards that they use to power the model.

Choi predicted that Mi:dm will generate 30 billion won ($22 million) in annual revenue domestically after three years. That number will grow if the model expands overseas.

“We target the B2B sector first because there is more demand from enterprises to find more cost-effective language models apart from the ones made by the Big Tech companies,” Choi said. “After three years, an annual revenue of 30 billion won in the local market, and even more possibilities when Mi:dm enters overseas.”

Earlier this year, KT announced that it aims to generate more than 1 trillion won in annual revenue through its AI business by 2025.

BY LEE JAE-LIM [lee.jaelim@joongang.co.kr]