Vietnamese telecom and technology group Viettel has unveiled VT-Super-120B-A12B, a 120-billion-parameter large language model (LLM) designed to strengthen the country’s sovereign artificial intelligence capabilities and support government and enterprise applications.
Developed by Viettel AI, the model is based on NVIDIA’s open Nemotron 3 Super architecture and has been further trained and optimized for Vietnamese language understanding, local knowledge, and domestic business use cases, according to the company.
Viettel said the model achieved performance levels among the leading AI models of similar size in internal evaluations, while improving Vietnamese-language capabilities without significantly reducing the original model’s English-language performance. The company also said it had mitigated “catastrophic forgetting,” a common issue in AI training where models lose previously acquired knowledge after additional training.
The company did not disclose benchmark scores, evaluation datasets, or comparisons with specific competing models.
VT-Super-120B-A12B was trained using Vietnamese-language datasets that include administrative records, business data, real-world conversations, and specialized documents, Viettel said. The training process involved continued pre-training, supervised fine-tuning, and reinforcement learning.
According to Viettel, the model is intended for applications in public administration, legal services, customer support, enterprise operations, and data analysis. Its long-context processing capabilities are expected to help organizations work with large volumes of documents, regulations, and complex workflows.
“AI is evolving from a supporting tool into a new layer of digital infrastructure,” Nguyen Manh Quy, Director of Viettel AI, said in a statement. “Countries that master their own models, data, and training capabilities will be better positioned to develop systems aligned with their national development goals.”
The announcement comes as governments and companies worldwide increasingly invest in so-called sovereign AI initiatives aimed at developing domestic AI infrastructure, models, and data ecosystems rather than relying entirely on foreign technology providers.
NVIDIA’s Director of AI Nations, Shilpa Kolhatkar, said Viettel’s efforts could help expand access to AI technologies while transforming local language and data resources into practical applications for government agencies and businesses.
Viettel said it is already building AI agents on top of the model, including an AI legal assistant designed to analyze documents, compare regulations, summarize information, and support decision-making processes.
While the company has provided details about the model’s architecture and training approach, technical information such as the size of the Vietnamese training corpus, the number of training tokens, benchmark results, inference costs, and hallucination rates has not yet been publicly released.
VT-Super-120B-A12B is expected to serve as a core AI platform for Vietnamese organizations seeking locally optimized AI systems with greater control over data, deployment, and regulatory compliance.
