Viettel AI Reports Initial Success in Joint LLM Development with NVIDIA

Viettel AI has partnered with NVIDIA to develop VAI-LLM, a 120-billion-parameter large language model built on the NVIDIA Nemotron 3 Super architecture. The model is designed to significantly enhance Vietnamese language comprehension and processing while maintaining core capabilities in reasoning, instruction following, and general knowledge.

To achieve this goal, the research team trained the model on massive, highly diverse datasets, ensuring strong performance in Vietnamese without compromising the foundational model's core strengths.

The development of VAI-LLM involves close collaboration between experts from both Viettel AI and NVIDIA. Throughout the pre-training, fine-tuning, and optimization stages, the team conducted large-scale experiments to streamline evaluation and maximize model performance.

Initial benchmark results indicate that VAI-LLM performs strongly across a wide range of tasks—spanning general knowledge, reasoning, and mathematics, as well as specialized Vietnamese language benchmarks. These early outcomes establish a solid foundation for further refining the model and evaluating its real-world application potential.

Evaluation Results of Viettel AI's LLM on the VMLU Fine-Tuned Benchmark

These initial results represent Viettel AI's first steps in its journey to build sovereign AI. However, this is only the beginning, and many challenges remain to be addressed in order to create practical, real-world AI solutions for Vietnam.

In the next phase, Viettel AI will continue refining its evaluation and real-world testing processes, while accelerating the development of specialized domain-specific AI applications built on top of the large language model.