The Institut bouddhique's rolling library, a truck racked with manuscript bundles, Phnom Penh, 1930

AI research and products for Khmer

Khmer speakers should be able to use AI in their own language. We’re doing the research to make that possible, and sharing every model and dataset we make, free.

Try it in your browser See how it’s built

“Rolling Library” of the Institut bouddhique de Phnom Penh, 1930 · public domain

Our models

Demo live

Netra-OCR

Turns photos and scans of Khmer text into text you can copy, search, and edit.

Params: 17M
Trained on: 1.3M images
Released: 2026-07-25
Report: Read it

Demo live

Netra-NMT

Translates between English and Khmer, in both directions.

Params: 77M
Trained on: 220M tokens
Released: 2026-07-11
Report: Read it

Demo soon

AkaraAlpha

Type Khmer in Latin letters, the way you'd text it, and get it back in Khmer script.

Params: 78.7K
Trained on: 28,569 samples
Released: 2026-06-15
Report: Read it

Read the full write-ups →

Why Khmer

More than 16 million people speak Khmer. Most AI still treats it as an afterthought.

There’s little public data to train on, and tools built for English stumble on a script where consonants stack and words run together without spaces.

We build for Khmer from the start, and publish everything, so the next person working on Khmer doesn’t have to start from zero.

Our commitments

Built for the people who read and write Khmer.

Read all our commitments →

You won't need an account, an API key, or a credit card to use anything we make. Code, weights, and reports are released under licenses that allow commercial use, changes, and redistribution.
This isn't a trial period. We won't lock the work away later, even if a business reason to do so comes up. Khmer-language tools shouldn't need anyone's permission to build on.
Read more →

No account, no API key, no paid tier. Everything we release is yours to use and build on.

You won't need an account, an API key, or a credit card to use anything we make. Code, weights, and reports are released under licenses that allow commercial use, changes, and redistribution.

This isn't a trial period. We won't lock the work away later, even if a business reason to do so comes up. Khmer-language tools shouldn't need anyone's permission to build on.

Every model is designed around how Khmer is written, not adapted from English afterward.

More than 16 million people speak Khmer, yet most AI tools barely understand it. Public training data is thin, test sets barely exist, and most systems treat Khmer as one more language added at the end.

The script itself breaks assumptions built into tools made for the Latin alphabet: consonants stack, vowels attach around them, and words run together without spaces. We start from how Khmer is actually written, and design each model around that.

Each release ships with its test data, its scores, and the places it still gets things wrong.

Every model comes with a written report: how it was trained, what it was measured against, and the test data we used, so you can check the numbers yourself.

We publish the failures alongside the scores. Knowing where a model struggles is more useful to the next person than a headline number, and you shouldn't have to take our word for anything.

The goal is a healthier Khmer AI ecosystem, not a moat around ours.

Netra Lab is independent. It's run by a single researcher, not a company, and isn't funded by anyone with a stake in the results. That leaves us free to put the work first.

We want a healthier ecosystem for Khmer-language AI, not a moat around our own models. If someone trains something better on top of this work, that's the point.