Viettel Trains 120-Billion-Parameter Vietnamese Sovereign AI Model, Ranking Among Leading Models of Comparable Scale

29/05/2026

Viettel AI has developed VT-Super-120B-A12B, a 120-billion-parameter Vietnamese large language model (LLM), with the goal of building AI systems deeply aligned with Vietnamese data, language, and domain knowledge.


Viettel Artificial Intelligence and Data Services Center (Viettel AI) is advancing sovereign AI development in Vietnam and demonstrating notable improvements in model capability following training and fine-tuning of NVIDIA Nemotron open models.


Through benchmarking evaluations, VT-Super-120B-A12B - the Vietnamese LLM researched and developed by Viettel AI – has demonstrated strong performance, and leads in accuracy among the models of similar size.


This achievement is the result of the model being trained on locally sourced data and optimized for domestic business applications. Notably, VT-Super-120B-A12B significantly enhances Vietnamese language processing capabilities without compromising the base model's English-language performance, while also mitigating catastrophic forgetting – a common issue in AI model training and fine-tuning where previously acquired knowledge is lost.


VT-Super-120B-A12B is built on the open NVIDIA Nemotron 3 Super architecture with 120 billion parameters, with training, fine-tuning, and Vietnamese optimization conducted by Viettel AI’s engineers. NVIDIA Nemotron’s long-context processing capability enables the model to preserve contextual continuity across multiple documents, workflows, and complex conversations within a single task.


Viettel Trains 120-Billion-Parameter Vietnamese Sovereign AI Model, Ranking Among Leading Models of Comparable Scale

VT-Super-120B-A12B is built on the open NVIDIA Nemotron 3 Super architecture with long-context processing capability (Photo: NVIDIA)


This capability is particularly critical for operational use cases in Vietnam, where regulations and procedures often vary depending on document types, stakeholders, and implementation contexts. Owning a core model capable of retaining contextual memory and adapting flexibly to real-world data provides the foundation for Viettel AI to fine-tune LLMs with high accuracy, enabling AI applications tailored to the specific operational needs of Vietnamese organizations and enterprises.


Beyond inheriting capabilities from the NVIDIA Nemotron 3 Super architecture, Viettel AI has also developed a dedicated LLM training pipeline built on localized data sources, including administrative data, enterprise operational data, real-world conversations, and specialized Vietnamese document systems. The process encompasses multiple training stages; in which the continued pre-training phase expands the model's Vietnamese linguistic capabilities and knowledge base across a large-scale data corpus. The model is then refined through supervised fine-tuning to improve reasoning performance, followed by reinforcement learning to enhance response accuracy and task execution capability.


Mr. Nguyen Manh Quy, Director of Viettel AI, stated: “AI is evolving from a supporting tool into a new layer of technological infrastructure. Countries that master models, data, and training capabilities will gain a strategic advantage in building systems tailored to their own development needs. For Viettel, mastering Vietnamese LLMs is a core step toward developing sovereign AI solutions capable of truly supporting Vietnamese organizations and enterprises.”


Viettel Trains 120-Billion-Parameter Vietnamese Sovereign AI Model, Ranking Among Leading Models of Comparable Scale

Mastering Vietnamese LLMs is a core step toward developing sovereign AI solutions for Vietnam.


Shilpa Kolhatkar, Head of Global AI Nations, NVIDIA shared: "By advancing sovereign, domain-specific AI, Viettel is helping democratize access to intelligence at population scale; turning local language and datasets into real world impact across government and enterprises."


Built on this foundation, Viettel AI is developing an AI Agent platform for Vietnamese users capable of autonomously executing multi-step workflows within a unified workspace. Among the first applications is the Legal AI Assistant, which is expected to support document analysis, regulatory cross-checking, data synthesis, and solution recommendations tailored to specific user scenarios, with significantly improved quality compared to previous versions.


From public administration and customer service to enterprise operations and data analytics, VT-Super-120B-A12B is positioned to become a core technology, powering AI systems that are designed to understand and address the real operational needs of Vietnamese organizations and enterprises.

Payment method
vnpay vtmoney
Banner_CTTDT_BQP2 Banner_CDVC_BQP2

logoSaleNoti