An independent lab for Khmer-language AI.
01/What this is
Netra Lab develops open-source models and tools for the Khmer language, each backed by a written report rather than a bare checkpoint and a claim. It is a small, independent effort — run by a single researcher, not a company, not a product, and not funded by anyone with a stake in the output.
Khmer is spoken by more than 16 million people, yet it sits far outside the languages that modern AI is built and benchmarked for. Public training corpora are thin. Evaluation sets barely exist. The script itself — with its stacked consonants, dependent vowels, and absence of spaces between words — breaks assumptions that are quietly baked into tooling designed for Latin alphabets.
That combination means Khmer speakers are usually handed systems that were built for another language and adjusted afterward. The lab exists to build the other kind: designed around the script from the start, and published openly enough that the next person doesn’t have to start from nothing.
02/The name
នេត្រ (netra) is the Khmer word for “eye.”
The logotype draws that word on a 29×24 pixel grid. That grid is not a stylistic flourish — it is how an OCR model reads a page: pixel by pixel, resolving structure out of an array of values.
The name comes from the Cat’s Eye Nebula, chosen for being beautiful, inspiring, and powerful in the way a nebula is, and for resolving into the shape of an eye.
The lab’s secondary imagery is the fishing cat, a small wild cat native to the region with some of the sharpest eyesight of any mammal — and a nod to the “cat” already hiding in Cat’s Eye Nebula.
03/Commitments
01
Ships as a report, not a checkpoint
Every model release comes with a written report: methodology, evaluation, and the baselines it was measured against. A checkpoint without that context is not a release, it's a claim.
02
Honest evaluation
Benchmarks are published with their evaluation sets, including the cases where the models do badly. Failure modes are more useful to the next researcher than headline numbers.
03
Permissive licensing
Work is released under licenses that allow commercial use, modification, and redistribution. Khmer-language tooling should not be something anyone has to ask permission to build on.
04
Free, always
No paid tier, no API key, no catch. Every model and report published here is free to use, fork, and build on — indefinitely, not just until there's a business reason to close it off.
05
Lightweight by design
Models are sized for the hardware people actually have, not a rented GPU cluster. A model nobody can run isn't accessible, whatever the license says.
06
Built to be replaced
The goal is a healthier Khmer-language AI ecosystem, not a moat. If someone trains something better on top of this work, that is the point.
04/How we work
Work happens in public on GitHub. Models are published to Hugging Face as they are finished, and each one gets a written report here: methodology, evaluation against existing systems, and the cases where it still falls short — not just a demo and a claim.
Netra-OCR, Netra-NMT, and AkaraAlpha are the results so far — purpose-built models, benchmarked and documented rather than announced. There is no roadmap to promise against: projects are picked because they unblock something concrete for Khmer-language computing, and because the research behind them is worth having in the open.