Open research lab · Bangkok · est. 2026

Good AI is AI we understand.

Like a bonsai, a model can only be shaped well by someone who knows how it grows. We are an open, non-commercial lab that studies how AI models are built, so they can be made smaller, fairer and safe to live with.

AI ที่ดีคือ AI ที่เราเข้าใจและควบคุมได้อย่างถ่องแท้ และอยู่กับเราได้ในชีวิตประจำวัน

SMALLER·SMARTER·CLOSER TO YOU

What we do

We read open models, prune what people don't need, and put them where people live.

  • 01
    Read

    Recover how a model is built from its files, without downloading its weights.

  • 02
    Prune

    Remove what a user's languages never use, and certify what the cut costs.

  • 03
    Deploy

    Run the result inside apps and on the phones people already carry.

Our four pillars

Four pillars, one question: can people understand, control and live with their AI?

  • L1
    Model Anatomy & Control

    Know what a model is made of before changing it.

  • L2
    On-device AI

    Make models small enough to carry, with proof nothing important was lost.

  • L3
    Human–AI Interaction

    How people use, trust and stop AI in daily life.

  • L4
    Multilingual Equity

    What each language pays, Southeast Asia first.

Scroll on to see the work inside each pillar.

L1 · Model Anatomy & Control

Every model grows like a tree you can read.

This is Qwen3-0.6B, drawn from its real config. The roots are its vocabulary. The trunk carries 28 rings, one per decoder block. The crown is the output head, which shares the roots' rows. The light climbing the trunk is a token passing through.

28decoder blocks1,024hidden width151,936vocabulary rows

L1 · The Hub census

We read 12,974 of them without downloading a single weight.

Configs and weight-file headers are enough to recover how a model is built. The whole Hugging Face Hub becomes a forest we can measure tree by tree.

12,974repositories0weights downloaded

Paper · Model Explorer

L2 · On-device AI · The vocabulary tax

Look underground: the roots outweigh any branch.

151,936 rows of 1,024 numbers each. That one matrix holds about as many parameters as ten decoder blocks, and a phone has to load all of it.

155.6Membedding parameters≈ 10×one decoder block

L2 · Pruning with a certificate

A Thai assistant never uses most of those roots. So we prune them, and certify the cut.

Roots for scripts a user will never type are cut away. Only the roots that feed the chip stay. Before anything ships, a certificate states what the cut cost in quality.

89.7%median rows removed where the certificate passes≤ +1.99%change in bits per character

Paper · Python package

L2 · On the device

Pruned well, it fits in your pocket.

A smaller model loads faster, uses less memory and runs inside an app or on a phone. What you say to it can stay on the device.

L3 · Human–AI Interaction

Living with it: AI you can talk to, trust and stop.

Our next study looks at voice assistants that run on the phone: how people speak to them, when they trust them, and how easily they can interrupt them. Thai speech in real rooms comes first.

FIRST STUDY PLANNED · 2027

L4 · Multilingual Equity

Every language grows its own leaves. Southeast Asia first.

Most models are built around English and priced in its tokens. We measure what each language pays and test every method on Southeast Asian languages first, then carry it to the rest of the world.

ไทยTiếng ViệtBahasa IndonesiaBahasa MelayuFilipinoខ្មែរລາວမြန်မာ+ the world

Tax table for 41 languages · Dataset

Projects

What we are working on

ACTIVE · L2 · L4

The Vocabulary Tax

Remove vocabulary rows a user's languages never need, with a certificate that bounds the cost before you ship.

ACTIVE · L1

Hub Census

The architecture of 12,974 public models, read from manifests alone and released as open data.

IN PROGRESS · L2

Slimming Gemma 4 E2B

Vocabulary and per-layer embedding trimming for a Thai, English and Chinese assistant on the phone.

PLANNED 2027 · L3

Voice assistants that stay on the phone

How people talk to, trust and interrupt an assistant that never sends audio off the device.

PLANNED 2027 · L4

SEA language tax table

A table that updates with every new open model: what Thai, Vietnamese, Khmer and others pay.

Publications

Papers with their data attached

2026
The Vocabulary Tax: Safely Trimming Unused Tokens from LLMs for On-Device UseSattaya Singkul · arXiv:2611.19913 [cs.CL] · L2 · L4
2026
Reading the Architecture off the Weights: Manifest-Only Model Anatomy and a Census of 12,974 Hugging Face RepositoriesSattaya Singkul · SSRN preprint · arXiv pending · L1

Roadmap

Three phases, two gates

PHASE 1 · NOWDec 2026 – Mar 2027

Open the lab

  • Lab site live with all outputs
  • Trimmed TH+EN models on Hugging Face
  • Compute grant applications
◆ Gate: one peer-reviewed paper
PHASE 22027

Build the name

  • Major-venue submissions
  • First yearly census report
  • SEA tax table, first L3 study
◆ Gate: 3+ outside collaborators, steady funding
PHASE 32028

Become an institute

  • Registered foundation
  • An international workshop
  • A yearly report others cite

Events & talks

Where to meet the lab

Get involved

Four ways to work with us

01

Contribute code

Pick an open issue on GitHub. Readers for new architectures are always welcome.

Open issues →
02

Bring a language

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

Add your language →
03

Research together

Universities, labs and model teams: co-author a study on our data.

Propose a study →
04

Support the lab

Compute, devices for on-phone testing, or a grant. Everything stays open.

Talk to us →

Or write to us: sattaya.singk@gmail.com

Collaborators & supporters

[ hardware partner ]
[ research partner ]
[ language partner ]
[ compute supporter ]

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.

FAQ

Is the lab a company?

No. It is an open, non-commercial research project. Nothing we publish is sold.

Can I use the data in my product?

Yes. The census and vocabulary-tax tables are MIT and the text corpus is Apache-2.0; each dataset states its license on Hugging Face. Please cite the paper you use.

Why Southeast Asia first?

Its languages pay some of the highest token costs and get the least testing. A method that holds here holds widely.

How do I cite the lab?

Cite the paper you used: arXiv:2611.19913 for the Vocabulary Tax, doi:10.2139/ssrn.7699619 for the Manifest census.

One email when we publish

New papers, tools and datasets. No more than once a month.