In the era of global AI explosion, technology products applying artificial intelligence are no longer the privilege of international technology corporations. In Vietnam, Zalo AI is gradually asserting its pioneering position by successfully developing a large language model (LLM) for Vietnamese, deploying a series of practical products. Typical of these are the Q&A assistant Kiki Info and the AI ​​greeting card application - two names that have quickly attracted millions of users within just two months of launch.

1. Vietnamese Big Language Model: A Proud Development Milestone

At the global level, LLM is no longer a stranger when technology giants such as OpenAI, Google or Meta continuously launch versions of GPT, Gemini or Llama with huge parameter scales, processing many languages. However, Vietnamese is a language with a smaller user community, limited training data and scarce digital language resources. Therefore, Zalo AI's self-development of a Vietnamese LLM model with a scale of up to 13 billion parameters has become an important step forward, affirming domestic technological capacity.

This model was built and trained entirely from scratch by Zalo's engineering team, deciding on the entire architecture, parameters, and optimization algorithms on a diverse Vietnamese data platform. This approach is highly appreciated by the technology community, because instead of relying solely on the open source model and refining it, Zalo chose the more arduous path of personally laying the first bricks for the Vietnamese LLM foundation.

2. Kiki Info Q&A Assistant

One of the outstanding achievements when applying this large language model is the Q&A Assistant Kiki Info. No longer just a theory in the research lab, Zalo's LLM has stepped into the real world in the form of an Official Account (OA) - an official account on the Zalo platform.

Kiki Info is designed as a general assistant, operating 24/7 to answer questions in many diverse fields: from science, history, law, traffic, to daily life knowledge. Anyone can ask questions and receive instant answers, with accuracy and semantics suitable for the characteristics of Vietnamese.

The difference of Kiki Info is not only in answering dry questions. Thanks to the well-trained Vietnamese LLM platform, Kiki Info also supports users in creating content such as composing emails, writing documents, creating social media posts, tasks that were previously time-consuming and required writing skills.

In addition, Kiki Info also provides entertainment and personalized suggestions, with recommendations on tourist destinations, music, and books that suit each user's needs. Thanks to that, this application quickly becomes a powerful support tool for work, study, and daily life.

A successful AI product is not only evaluated by technology but also by its ability to reach the community. Statistics from the development team show that Kiki Info has reached more than 1 million visits in just under two months, becoming one of the fastest growing OAs on the Zalo ecosystem.

This proves that the need for a friendly and accessible Vietnamese virtual assistant is huge. In the context of global AI still having many language barriers with Vietnamese, Kiki Info has demonstrated the development potential of domestic AI products - when technology is truly linked to the needs and habits of Vietnamese people.

3. AI Card

Along with Kiki Info, Zalo also launched a unique application: AI Cards. Not just a simple technology product, AI Cards also have the meaning of connecting people, bringing a new breeze to the culture of sending greetings on holidays, Tet, birthdays or important events.

In just two months, the AI ​​greeting card application has reached the milestone of 15 million cards created and sent by users. This number not only reflects the attraction of technology, but also shows that Vietnamese people are ready to accept new, personalized experiences when AI technology becomes more friendly and closer than ever.

Users just need to enter the desired content, choose the style, AI cards will automatically create interesting wishes, eye-catching designs to send to friends and relatives. In the busy life, this is a subtle way to maintain connection and spread care in a modern way.

4. Practical benefits

The plus point of Kiki Info and AI cards is not only in the underlying technology but also in the way they solve very practical problems. With Kiki Info, users do not have to waste time looking up and searching for information from many different sources but can receive instant, accurate, and easy-to-understand answers. Similarly, AI cards help save effort in thinking of good wishes, designing beautiful cards and sending them in just a few steps.

These applications are highly appreciated by the user community for their convenience and scalability, especially the personalization suitable for the Vietnamese context - something that many international AI assistants have not really done well. Thanks to that, Zalo's AI products have contributed to shortening the time to search for information, making decisions faster and maintaining connections between people in a creative and humane way.

Behind the impressive numbers and useful products is a challenging journey for the Zalo AI team. Compared to international giants like OpenAI or Meta, Vietnamese businesses face many barriers in terms of hardware infrastructure and data resources.

From the very beginning, while global companies owned Nvidia’s modern GPU systems, server infrastructure in Vietnam was limited in both quantity and quality. Not to mention, Vietnamese is a language with much poorer data resources than English or Chinese – which are favored in large AI projects. Human resources were also not really abundant, lacking experience in large-scale LLM training.

5. Build your own training data

Another difficult problem is the Vietnamese data source. While English has countless public, standardized data sets for LLM training, Vietnamese is still fragmented, inconsistent, and unsystematic.

To overcome this, the Zalo engineering team proactively collected, cleaned and developed a suitable Vietnamese dataset for LLM training. Before owning the DGX H100 server system, the development team also took the opportunity to run tests on smaller civilian GPUs, to hone their experience, optimize the algorithm and be ready to "race" when there is a large infrastructure.

This is a way to demonstrate the determination to develop domestic technology systematically, without depending on the outside. It is these silent efforts that have laid an important foundation for Zalo to confidently launch Vietnamese LLM, on par with many global names.

6. Impressive milestones

The results have proven the right and persistent choice. At the end of 2024, Zalo's LLM model reached the top 2 position on the VMLU ranking - a platform for evaluating and ranking Vietnamese language processing capacity for LLM models.

More notably, Zalo has surpassed many big names such as OpenAI's GPT-4, Google's gemma-2-9b-it or Microsoft's microsoft/Phi-3-small-128k-instruct, only behind Meta's Llama-3-70B - a "giant" with a huge infrastructure and global data ecosystem.

Not stopping there, in 2023, Zalo announced a 7 billion parameter model focusing on Vietnamese, after only 6 months of training, it achieved 150% capacity compared to GPT-3.5 (according to VMLU standards). This shows that the domestic engineering team can confidently create high-quality AI products, not inferior to the world.

The initial successes with Kiki Info, AI cards and the Vietnamese LLM model are just the first bricks. Zalo representatives affirmed that they will continue to invest in development to take the LLM model further, opening up many new AI applications to serve the needs of Vietnamese people.

This direction is completely consistent with the national orientation on science and technology, innovation and digital transformation, which are three key factors for Vietnam not to be left behind in the global technology race. In the context that AI is considered the “key” to opening a new generation of products for all industries, mastering the large native language model will become a long-term strategic advantage.

In the near future, Vietnamese users will certainly welcome many new, smarter, and more convenient AI applications from this large language model. And who knows, today's pioneering steps will lay the foundation for Vietnam to become a "bright spot" on the global AI map.