About the lab

Understand AI from the inside, before we use it.

AI Model Explorer Lab is an open, non-commercial research lab in Bangkok. We study how AI models work inside, then use what we learn to build versions that run on limited hardware, that people can control, and that Thai and other Southeast Asian languages can afford.

Founded 2026Bangkok, ThailandOpen and non-commercial2 papers · 3 datasets · 1 tool

Why we started

Thailand uses a lot of AI. Few of us study what is inside it.

Most AI work in Thailand starts from the use case: pick a model, plug it in, ship. That is useful, but it leaves a gap. Few people here study how the models themselves are built, so it is hard to say what a model needs, what it wastes, or where our data goes when we use it.

We think that understanding should come first. Once you can read a model from the inside, you can make it smaller for a phone, control what it does, design how people work with it, and run it yourself, without sending data out of the country.

  • Understanding

    AI is used everywhere, but few people study what is inside the models.

  • Data leaves the country

    Most AI runs on servers abroad, so Thai data has to travel to use it.

  • Thai pays more

    In most models the same text costs more tokens in Thai than in English, and Thai gets less testing.

  • No place to work together

    Deep AI research here is scattered across people and teams that rarely meet.

Research like this is rare in Thailand, and the people who do it often work alone. The lab is a place for them to work together on problems Thai researchers can solve, and to take the best results to journals and conference stages.

Vision and goal

Good AI is AI we understand, can control, and can live with every day.

  • VISION

    AI people understand

    Before a model is shrunk, deployed or trusted, someone should be able to say what it is made of and what each part does.

  • GOAL

    AI Thailand can run itself

    Models that run on our own phones, laptops and servers, so data never has to leave the country.

  • OUR PLACE

    One layer of sovereign AI

    National programmes build compute and large Thai models. We work on the layer that reaches people: open models, understood and rebuilt to run on the hardware people already have.

Four pillars

Four pillars, one main theme.

Our main theme: understand AI models from the inside, so they can be made smaller, controlled, used well by people and fair to every language. The four pillars are the four parts of that theme. Each covers a broad area; the results under "Where we are now" are where we started, not where we stop.

  1. Model Anatomy & Control
  2. On-device AI
  3. Human–AI Interaction
  4. Multilingual Equity

The bonsai behind this page is drawn from a real model: its roots are the vocabulary, the trunk its 28 blocks, the crown its output. As you scroll, the camera moves to the part each pillar works on.

L1 · Read the inside of a model

Model Anatomy & Control

What is inside a model, and which parts can be changed safely?

How AI models are built, and how to steer them. Before anything is changed, we want to know what every part of a model is and what it does.

Why it matters. You cannot shrink, fix or control what you cannot see. Most people choose models by name and benchmark score, not by what they are made of.

WHAT THIS PILLAR COVERS

  • Reading architecture from model files
  • What layers, heads and embeddings do
  • Tokenizers and vocabulary structure
  • Editing and steering model behaviour
  • Trends across thousands of public models
  • Model reports anyone can reproduce

WHERE WE ARE NOW

12,974models read, no weights downloaded

DONE SO FAR
Manifest-only anatomy reads a model's architecture from its config and weight-file headers. With it we mapped 12,974 Hugging Face repositories without downloading a single weight, and built AI Model Explorer so anyone can do the same.
NEXT
Readers for more model families and a yearly census report.

L2 · Make it small and sturdy

On-device AI

How small can a model get and still do its job?

Making AI run well on the devices people own: phones, laptops, boards and servers inside the country, with a measured cost for every change.

Why it matters. A model that only runs in a data centre abroad cannot keep data in the country. Phones and laptops have little memory, so a model has to be rebuilt for them, not only squeezed.

WHAT THIS PILLAR COVERS

  • Pruning, quantization and distillation
  • Vocabulary and embedding trimming
  • Small multimodal models: text, voice, images
  • Robustness under low memory and offline use
  • Benchmarks for speed, memory and energy
  • Running models inside apps and on edge chips

WHERE WE ARE NOW

89.7%median vocabulary rows removed, quality certified

DONE SO FAR
The Vocabulary Tax removes vocabulary rows a user's languages never need and certifies the cost first: a median of 89.7% of rows removed where the certificate passes, with bits per character changing by at most +1.99%.
NEXT
An optimized multimodal model that runs entirely inside an app, and tests of how models hold up under low memory, low-bit weights and offline use.

L3 · Make it work for people

Human–AI Interaction

How do people use, trust, correct and stop AI that runs on their own device?

How people live with AI that runs on their own devices: how they use it, how much they trust it, and whether they stay in control.

Why it matters. A model that runs well but that people misread, over-trust or cannot stop is not good AI. How people work with it decides whether it helps.

WHAT THIS PILLAR COVERS

  • Trust and over-reliance
  • Correcting and stopping AI
  • Interaction through text, voice and images
  • Explanations people can act on
  • Privacy and on-device data
  • User studies in Thai contexts

WHERE WE ARE NOW

2027first study, with Thai users

DONE SO FAR
Nothing published yet. This pillar starts in 2027.
NEXT
A first study with Thai users, through text, voice and images.

L4 · Make it fair for every language

Multilingual Equity

What does each language pay in cost and quality?

Making AI fair across languages, so a Thai or Khmer speaker gets the same quality at the same cost as an English speaker.

Why it matters. Most models are built around English. Thai and other Southeast Asian languages need more tokens for the same text, so they cost more and run slower.

WHAT THIS PILLAR COVERS

  • Token cost per language
  • Quality gaps between languages
  • Test sets for under-tested languages
  • Trimming models to the languages people use
  • Southeast Asian language coverage
  • Gaps in data and benchmarks

WHERE WE ARE NOW

41languages in the open tax table

DONE SO FAR
A vocabulary tax table for 41 languages, released as open data.
NEXT
A Southeast Asian tax table that updates with every new open model.

How we work

Four rules every output follows

  • Open by default

    Code, data and the numbers behind every claim.

  • Runs on a laptop

    Core methods repeat on ordinary hardware.

  • Proved before claimed

    Each change ships with a measured cost bound.

  • Built from Southeast Asia

    Non-English is the first test, never an appendix.

Everyone who makes a substantial contribution is an author, and author order is agreed when a project starts. How to join as a researcher →

Founder

Who started the lab

SS
FOUNDER · LEAD RESEARCHER

Sattaya Singkul

Jo · Bangkok, Thailand

Jo works on speech AI and on-device deployment. He wrote the Manifest census of 12,974 Hugging Face repositories and The Vocabulary Tax, and built AI Model Explorer for Thai developers. He started the lab so that this work keeps growing in the open.

Who we work with

The lab grows through the people and teams who join it.

  • Researchers and students

    People in Thailand who want to do deep AI research and publish it.

  • Universities and institutes

    Co-supervised studies, shared data and joint papers.

  • Thai model teams

    We would like to read, test and rebuild your models with you, for the devices people use.

  • Hardware makers

    Phones, boards and laptops for testing models on the device.

  • Funders and supporters

    Compute, publication and travel costs. Everything we make stays open.

  • Language communities

    Native speakers who build test sets for languages the big benchmarks skip.

We name partners here once we have agreed to work together.