The explosion of Generative AI has completely changed the global technology landscape.
Since OpenAI launched ChatGPT in late 2022, a series of similar LLM models have continuously appeared, applied in many fields such as healthcare,education , finance, law, etc. The race has become fierce, not only at the enterprise scale but also at the national scale in terms of training capacity, computing infrastructure and data.
Large companies such as OpenAI, Google, Meta or Microsoft with financial potential and favorable conditions have all "quickly" invested billions of dollars to own thousands of high-end GPUs such as Nvidia H100 - the core factor for training LLM models. Nvidia's stock price skyrocketed during that period, reflecting the world 's thirst for infrastructure.
Meanwhile, technology companies in developing countries, in addition to cost issues, also face restrictions on importing and exporting AI chip equipment from the US. This leads to a lack of training equipment and being "slower" compared to technology giants. With domestic aspirations, Zalo is no exception.
In terms of data, previous machine learning problems required large amounts of training data, but for large language models, this is a huge amount of data. To have a good enough model, the LLM system requires tens or even hundreds of billions of input text tokens. Meanwhile, in terms of popularity, Vietnamese is not on par with English and Chinese. This has multiplied the difficulty for Vietnamese LLM developers.
In 2023, large language models (LLM) such as GPT-3.5, GPT-4 have taken the technology world by storm, and many Vietnamese enterprises have also chosen foreign-made fine-tuned models to take a shortcut and get ahead of the LLM training process.
Zalo chose a different path - more arduous, but more autonomous: Self-training the model from scratch (from-scratch model). A path that requires everything to be built from scratch - from data, model architecture to the entire training process. That decision was not to compete with the giants but to realize the aspiration of the Vietnamese people themselves: mastering the LLM model in their mother tongue.
“We anticipated the difficulties and still decided to join the game early. We did not compete directly with the “big guys” but chose a niche market where we could do better. Our aspiration is to build a model that Vietnamese people completely control - from data to algorithms,” shared Dr. Nguyen Truong Son, Chief Science Officer at Zalo AI.
Despite facing many obstacles in three main factors including infrastructure, data and training level, Vietnamese engineers have proactively found solutions to overcome difficulties. This demonstrates the spirit and willpower of the Vietnamese people in difficult circumstances, specifically in this case, conquering challenges in the process of developing LLM for Vietnamese people.
To train LLM, engineers needed the right infrastructure. But at that time, GPUs like Nvidia’s H100 were almost “global rarities”. Meanwhile, big companies had pre-ordered them for a year and paid millions of dollars to own them. In Vietnam, Zalo also tried to buy 8 DGX H100 servers but it was not easy, having to wait for each batch of deliveries from the manufacturer.
In the absence of Nvidia GPUs, Vietnamese engineers had to flexibly use civilian GPUs to experiment on each line of code and run each small model. Instead of waiting, engineers proactively prepared so that when they had modern equipment, everything would be ready.
In terms of data, instead of relying on available resources, Zalo invested in building a high-quality data warehouse specifically for Vietnamese, to make up for the serious shortage compared to English and Chinese.
Thanks to its flexible development strategy, Zalo has successfully shortened the development time of its large language model from the expected 18 months to 6 months. At the end of 2023, Zalo's Vietnamese large language model was officially launched at the event that gathered the leading technology and AI community in Vietnam - Zalo AI Summit. Here, Zalo's LLM model made its debut through the Kahoot challenge set by Tinhte.vn and surprisingly surpassed GPT 3.5, only behind GPT4 - the LLM model that was considered the strongest in the world at that time.
On the VMLU (Vietnamese Multitask Language Understanding Benchmark Suite for Large Language Models) evaluation platform, Zalo's model achieved 1.5 times the capacity of OpenAI's GPT-3.5. By the end of 2024, this model will surpass big names such as GPT-4 (OpenAI), Gemma-2-9B (Google) or Phi-3-small (Microsoft), only behind Meta's LLaMA-3-70B in terms of Vietnamese processing capacity on VMLU's rankings.
Not only stopping at research, Zalo is gradually bringing technology from the laboratory to life by commercializing and popularizing application products from LLM.
In early 2025, the Q&A assistant Kiki Info - operated as an official account on the Zalo platform - attracted more than 1 million users in less than 2 months. Another application, Thiep AI, also reached an impressive number of 15 million cards created and sent in just 2 months.
Zalo’s journey is not just about a company wanting to develop technology. It is a piece of the bigger picture - where Vietnam is aggressively promoting innovation, with policies from Resolution 57-NQ/TW on science, technology development and national digital transformation. In particular, the field of artificial intelligence is emphasized.
The emergence and rapid development of Vietnamese LLM from Zalo is not only a technological step forward for a business, but also a testament to the inherent capacity and perseverance of the Vietnamese technology team.
With the “from-scratch” technique - training models from scratch, Zalo chose the long road, but helped Vietnam truly master AI. Not only in terms of results, but also in terms of the entire process from model architecture, data, algorithms, to application products. Zalo's success has also helped Vietnam become one of the few Southeast Asian countries to own a domestic LLM model - a strategic milestone in the context of increasingly fierce global technology competition.
On the long journey ahead, Zalo will not only stop at one model or a few products but will continue to perfect the model to both serve users and create a competitive Vietnamese AI platform: “Zalo's AI development journey is still long. We will continue to optimize the model in both breadth and depth, while promoting practical application. The ultimate goal is to create quality AI products that practically serve Vietnamese people," Mr. Son added.
Zalo’s successful development of Vietnamese LLM is not only a breakthrough for a business, but also opens up a potential future for Vietnamese artificial intelligence. The perseverance and aspiration of the Vietnamese people have led the journey to reach worthy results. The future of Vietnamese AI will not only have a “Zalo”, but also a generation of brave engineers to follow, inherit and conquer the world of technology.
Source: https://znews.vn/zalo-va-hanh-trinh-lam-chu-llm-tieng-viet-post1561765.html
Comment (0)