L1.0: AIVISION's Vietnamese LLM and What It Can Do

01/10/2026

A language model answering fluently does not mean it is right. That sounds obvious, but it is the most expensive lesson anyone building AI products learns at least once. When we at AIVISION released L1.0, a Vietnamese LLM, we wanted to introduce it in exactly that spirit: say what it is, where it is useful, and where it should not yet be trusted.

What L1.0 is

L1.0 is a large language model for Vietnamese, trained by AIVISION on a GPU cluster of 24 NVIDIA H200 and 8 NVIDIA B300 GPUs. The L stands for language, just as the E in E1.0 points to the speech recognition side, and the 1.0 is a reminder that this is a first version.

You will not find parameter counts, training data size, benchmark scores or rankings here. We have not published them, and we do not want to guess or round them up to look good. AIVISION's view is that a model should be judged by using it on your real work, not by a number in an announcement.

What a Vietnamese LLM can do

This section describes the typical abilities of a large language model working in Vietnamese, as a general picture of the kind of work you can hand to such a tool. How well each task goes depends on the specific model and how you use it, so test before you trust.

• Drafting and editing text: first drafts of emails, announcements, posts, then adjusting tone to be more formal or friendlier.
• Summarizing: condensing long documents, meeting minutes or message threads into a few key points.
• Answering questions from documents: a user asks, the system searches an internal document store and restates the answer in Vietnamese.
• Extracting information: pulling dates, names, contract numbers and key clauses from unstructured text.
• Classifying and labeling: sorting customer feedback, routing requests into groups.
• Reformatting: spoken text to written text, paragraphs to tables, scattered notes to a report.

Combined with a speech-to-text model, this chain runs the full loop: speech becomes text, text becomes a summary. That is why E1.0 and L1.0 are developed side by side.

Why Vietnamese needs its own model

Large multilingual models can handle Vietnamese fairly well, but we still see gaps. Forms of address are one example: anh, chi, em, ong, ba, chau, each choice carries a relationship, an age, a level of politeness. Pick the wrong pair and a grammatically correct text still sounds like a foreigner wrote it. The same goes for idioms, euphemisms, the particular style of administrative writing, and word choices that vary by region.

We focus on these details because for Vietnamese users, they are the difference between a tool that is usable and one that is pleasant to use. But to be honest, we do not claim to have solved it all. There is a lot left to do, and each later version is a gradual step rather than a leap.

Finetuning for domains

A base model is only a starting point. Practical value often comes from LLM finetuning: continuing training on data from a specific field so the model becomes familiar with its vocabulary, style and procedures. AIVISION does this for Vietnamese and for fields such as healthcare and pharma.

For healthcare and pharma we draw the line clearly. The model is only a tool for information and administrative support, for example summarizing documents, drafting, and organizing information for a qualified person to review. It does not diagnose, prescribe or advise on dosage. The medical expertise of doctors and pharmacists always has the final word. The blunt reason: an LLM can invent something that sounds very convincing, and in medicine, convincing but wrong is the most dangerous kind.

Limits you should know

Every large language model, L1.0 included, shares the general limits of this technology.

• It can make things up. When it does not know, a model sometimes still answers with confidence. For important documents, check the source.
• Knowledge has a cutoff. The model only knows what was in its training data, unless it is connected to an up-to-date source.
• It depends on how you ask. The same request phrased differently can give different results.
• It does not replace human decisions where legal, financial or medical consequences are involved.

We treat stating these limits as part of the product. Users who know what to trust and what to verify use the tool far more effectively than those who were promised too much.

How to test an LLM for your own work

A practical tip: do not test with vague prompts like describe Hanoi. Take five to ten real situations from your job, along with answers you consider correct, and see how the model handles them. Pay attention to the failures, because the kind of failure matters more than the rate: clumsy mistakes are easy to catch, fluent ones are what you should worry about.

If you follow Vietnam AI and want to watch how L1.0 develops, visit the AIVISION homepage. With the same approach, we also look at markets like Mexico and the Philippines, where users need models that understand local language and culture instead of merely translating from English.