40% Fewer Parameters... Operating Costs Reduced by Half
Naver has completed the update of its HyperCLOVA X flagship model, which emphasizes low cost and high performance, and unveiled it internally on the 20th.
The new HyperCLOVA X outperformed the previous model in average scores across all fields when comparing performance on 19 benchmarks for major training data including Korean, English, coding, and mathematics. In the large-scale multitask language understanding (MMLU) test, which evaluates knowledge and problem-solving abilities across 57 subjects such as mathematics, physics, history, law, and medicine, it achieved an accuracy rate of 79.6%. Its multi-modality capabilities have also been enhanced, adding functions that understand not only text but also images and videos.
Despite these significantly improved performances, the new HyperCLOVA X model is characterized by a cost-efficient structure. The number of parameters used in this model was only about 40% of the previous one. Additionally, the operating costs of the new HyperCLOVA X model have improved by more than 50% compared to the previous model.
Naver plans to integrate the new HyperCLOVA X model into its conversational AI service 'CLOVA X' next month. It also plans to release the new model on 'CLOVA Studio,' Naver Cloud Platform’s hyperscale AI development tool, to provide it to enterprise customers. In the second half of this year, a HyperCLOVA X model capable of natural voice conversations is scheduled to be unveiled.
Seong Nak-ho, Head of Hyperscale AI Technology at Naver Cloud, said, "In a situation where low-cost, high-performance AI models are gaining attention, Naver has also been steadily researching and developing such technologies," adding, "We will continuously enhance the capabilities of our core models, such as improving inference abilities and expanding modalities, to build AI technologies that can compete globally."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
