On 6/7, the AI Research Center at VinUni University unveiled V-Bench, a non-profit toolkit for assessing the Vietnamese language capabilities of large language models (LLMs). The tool was developed to support the research community, businesses, and organizations applying AI in Vietnam.

V-Bench was developed with the participation of an 18-member scientific council comprising Vietnamese AI experts from both domestic and international institutions. Its dataset includes over 40,000 questions and tasks, covering areas such as language, culture, history, digital sovereignty, information security, and various application contexts specific to Vietnam.

AI tool for Vietnamese language proficiency testing platform developed by VinUni. *Photo: VinUni*

The evaluation system features two categories with five criteria groups: implicit culture, regional diversity, safety and digital sovereignty, practical application, and agentic AI capability. These criteria assess a model's ability to comprehend language, dialects, cultural context, meet information security requirements, and its capacity for planning, tool use, and task execution within the Vietnamese environment.

According to VinUni, V-Bench aims to provide an independent reference system for evaluating AI models, offering organizations and businesses a basis to select models suited to their needs. The toolkit was developed using common international LLM research methods, ensuring compatibility in processes and terminology with global benchmark systems.

Professor Duong Nguyen Vu, Vice Rector and Science Director of the AI Research Center at VinUni University, stated that with its comprehensive dataset and the objective, non-profit contributions from reputable Vietnamese experts globally, V-Bench seeks to redefine AI standards in Vietnam and build a national AI reference system.

"V-Bench's evaluation results will offer a comprehensive view, helping organizations and businesses explore the depth of each AI model's capabilities, thereby enabling them to choose the optimal model for specific usage contexts and practical needs," he added.

V-Bench compares five artificial intelligence models. *Photo: VinUni*

To date, V-Bench has published reference evaluation results for 15 LLM models. In its next phase, the development team plans to add evaluation criteria related to images and audio, including Nom script recognition, regional signs, charts, cultural heritage, voices from three regions (North, Central, South), cultural context via short videos, and the ability to process long Vietnamese documents over 100,000 tokens, such as laws, contracts, textbooks, and legal texts.

Nhat Le

V-Bench is available for free at https://vbench.ai/.