Netra Lab is a small, independent research lab building open-source, lightweight AI for the Khmer language — OCR, translation, language models, and whatever the language needs next. Everything is released free, for anyone to use.
01/Why
Khmer is spoken by more than 16 million people, but the language is still underserved by modern AI. Training data is scarce, tooling is scattered, and most existing systems treat Khmer as an afterthought bolted onto models built for other languages.
Netra Lab exists to close that gap: building models, datasets, and tools specifically for Khmer, from the ground up, and releasing all of it publicly rather than keeping it behind an API or a paywall.
This is ongoing work, not a finished product — some of it is early, and it will keep evolving as the research does.
02/Principles
01
Free, always
No paid tier, no API key, no catch. Every model, dataset, and training recipe we publish is free to use, fork, and build on — indefinitely, not just until there's a business reason to close it off.
02
Built for Khmer, not adapted to it
Khmer orthography stacks consonants, uses no spaces between words, and breaks assumptions baked into tooling designed for Latin scripts. We design around the script rather than bolting it onto systems built for other languages.
03
Small enough to actually run
Lightweight by design: models sized for the hardware people actually have, not a rented GPU cluster. A model nobody can run isn't accessible, whatever the license says.
04
Reproducible or it didn't happen
We publish training recipes, evaluation sets, and benchmark numbers alongside the model, so anyone can verify the claims or find where they fall short.
03/What we publish
Code
Full training and inference source, released under permissive licenses so anyone can build on it.
Research
A written report for every model — methodology, evaluation, and benchmarks against existing systems, including where it falls short. Not a claim without the work behind it.
Weights
Trained model checkpoints, hosted on Hugging Face and free to download, fine-tune, or deploy.
04/Latest
Netra-OCR
A 17M-parameter encoder-decoder model trained from scratch on 1.3M bilingual Khmer/English text-line images, designed for high accuracy at low latency.
Netra-NMT
A 77M-parameter translation model trained from scratch on 220 million tokens of English–Khmer parallel text, using a deep-encoder / shallow-decoder layout for low-latency decoding.
AkaraAlpha
A 78.7K-parameter attention-based bidirectional GRU that converts romanized Khmer written in the Latin alphabet back into Khmer script, with dictionary-based post-processing.
Everything we build is public.
Use the models, fork the code, retrain on your own data, or tell us where they fall short.