Naver announced on the 20th that it has completed the update of the HyperCLOVA X flagship model and publicly unveiled it within the company. Using the newly unveiled HyperCLOVA X foundation model, Naver has also begun developing various artificial intelligence (AI) services for users, sellers, creators, and business partners.
The updated HyperCLOVA X model is composed of a relatively small-sized model with about 40% of the parameters compared to the existing model, but features stronger performance. A comprehensive performance comparison using 19 benchmarks for the key training data in Korean, English, and coding/math showed that the average score in all fields surpassed that of the previous model. Notably, in the globally recognized benchmark 'MMLU (Massive Multitask Language Understanding),' it recorded an accuracy of 79.6%, demonstrating language understanding capabilities comparable to similar-sized overseas big tech AI models.
"Multi-modality" capabilities have also been enhanced. The existing model's ability to process visual question-and-answer tasks and understand charts and diagrams, which could handle not only text but also image data simultaneously, has been improved to the performance level of the world's top models. New functionality for understanding video, beyond images, has also been added.
Additionally, the model was designed with an efficient structure to reduce operating expenses. According to Naver, the operating expense of the new HyperCLOVA X model has improved by more than 50% compared to the previous model. Under the 'On-Service AI' strategy announced last year, Naver, which is incorporating generative AI technology into major services with large user bases, such as search and commerce, is expected to accelerate AI application expansion with the newly low-cost, high-performance HyperCLOVA X model.
Since unveiling HyperCLOVA X in August 2023, Naver has continuously evolved the model to keep pace with global AI technology trends. In April 2024, it released the lightweight model 'HyperCLOVA X DASH' at a price reduced to one-fifth of the original, lowering barriers to generative AI adoption for corporations. Subsequently, in August, it unveiled the 'HyperCLOVA X Vision' model, which can process both text and images simultaneously.
Furthermore, it is developing technology that enhances planning and reasoning capabilities to systematically and comprehensively perform tasks requested by users, and plans to unveil a HyperCLOVA X model capable of natural voice conversation in the second half of the year.
Sung Nak-ho, head of Naver Cloud's hyper-scale AI technology, noted, "Recently, technologies that enable the operation of high-performance AI models at low expense have been gaining attention, and as Naver aims to reliably integrate AI into services used daily by millions, we have consistently researched and developed such technologies." He added, "I hope the new HyperCLOVA X model will serve as an engine providing differentiated AI service experiences to more users, and we will continue to enhance the capabilities of our flagship models, such as improving reasoning abilities and expanding modalities, to establish AI technologies that can compete at a global level."
Naver plans to equip the upgraded HyperCLOVA X foundation model in its conversational AI service Clova X in March so that users can experience its capabilities more intuitively. Additionally, there are plans to release it through Naver Cloud's hyper-scale AI development tool 'Clova Studio' for corporate customers.