The application of Vietnamese AI model is attracting millions of visits.

The application of the Vietnamese AI model is attracting millions of visits.

Developed by Zalo's Vietnamese engineering team, the large-scale language model with 13 billion parameters is driving a series of practical applications that attract a large number of users monthly. A standout is the Kiki Info comprehensive Q&A assistant, operating as an Official Account (OA) on Zalo, offering three features: Q&A on various topics such as science, history, traffic laws, etc.; Content creation such as writing essays, composing emails, and posting on social media; and Entertainment by providing suggestions on travel destinations, music , books, etc.

Statistics from the development team show that this assistant has had up to 1 million users accessing their Zalo Official Account (OA) in just under 2 months.

Kiki Info is integrated into the Zalo messaging platform.

Another application developed using Zalo's large language model is AI greeting cards, which have now reached 15 million cards created and sent by users. This data was compiled over a two-month period, showing users using AI greeting cards to send wishes to family and friends on important occasions.

These apps are highly rated by users for their intelligent experience, which helps shorten search times and decision-making in life. They also make connecting with others and building relationships more enjoyable and engaging.

Zalo's LLM model is developed using a cross-training technique – implementing all processes from parameter initialization and model architecture decision-making to training algorithms on a given dataset, allowing Vietnamese users to fully master and control the training process and the model.

At the end of 2024, Zalo's LLM model finished in the Top 2 position on the VMLU ranking – a platform for evaluating and ranking the Vietnamese language proficiency of LLMs (Vietnamese Multitask Language Understanding Benchmark Suite for Large Language Models). Specifically, Zalo's large language model rose to the number 2 position in the ranking of models trained from scratch, only behind Meta's Llama-3-70B, officially surpassing major names such as GPT-4 (OpenAI), gemma-2-9b-it (Google), and microsoft/Phi-3-small-128k-instruct (Microsoft).

The 2024 ranking of LLMs built from scratch, with Zalo's LLM model in the Top 2.

This was a huge success for a large-scale programming language developed by Vietnamese people, especially considering the numerous limitations it faced in its early stages. While large companies worldwide owned thousands of the latest GPUs from Nvidia, in Vietnam at that time, engineers were not yet equipped with the necessary server infrastructure.

At the same time, Vietnamese is also ranked among languages with data resources that are dozens of times poorer than English or Chinese. In addition, Vietnam also has limitations in human resources and experience in training LLMs when compared to developed countries around the world.

Zalo had a development strategy to overcome the limitations of its training environment compared to the rest of the world. By equipping its computing infrastructure with 8 DGX H100 servers, the LLM model was developed directly using Nvidia's newest and rarest GPUs at the time, with performance up to 256 petaFLOPS (Floating-point Operations Per Second - one petaFLOP is equivalent to 10 quadrillion calculations per second).

Zalo's server system has superior processing capabilities.

Simultaneously, investment in developing high-quality training data has been made to compensate for the shortage of Vietnamese-language data sources. Through a series of studies conducted on small consumer GPUs, Zalo engineers have also taken advantage of opportunities to acquire knowledge and training capabilities in LLM, creating a foundation ready for use when they acquire large-scale computing infrastructure.

The right development strategy has helped Zalo successfully develop a large-scale language model with 7 billion parameters focusing on Vietnamese in just 6 months of training in 2023, achieving 150% of the capability compared to OpenAI's GPT3.5 on the VMLU benchmark. Currently, it has surpassed numerous global names on the VMLU 2024 ranking and is bringing its research model into practical implementation for the community.

Zalo's LLM model competed against a range of global models such as ChatGPT 3.5, ChatGPT 4.0, Llama, PhoGPT, and a real player when it first launched in 2023.

According to Zalo, the LLM model will continue to receive investment in training to bring more world-class AI applications to users. This will enable them to master advanced AI technology, moving towards a new era of technological development for the country with a breakthrough orientation in science , technology, innovation, and national digital transformation.

Comment (0)