What’s that glyph?
Thai does not use Latin glyphs. That breaks Latin-first assumptions in tokenizers, datasets, OCR, and everything built on them.
The frontier is built around a handful of languages. Ours isn't one of them.
Wayu is a non-profit AI research initiative in Bangkok, founded by the creator of Typhoon. We work out why models misread, mishear, and mispronounce Thai — and share every understanding, every answer, so any language that fails the same way can compete.
Thai is treated as an edge case. We start here because we are Thai — and the same failures can extend far beyond it.
Thai does not use Latin glyphs. That breaks Latin-first assumptions in tokenizers, datasets, OCR, and everything built on them.
In Common Crawl, Thai is 0.4% of the web. 30% of it is casino sites — not the web Thai people use, but the data models learn from.
Everyday Thai mixes Thai and a lot-of-things. Models trained on textbook Thai learn a language nobody uses.
Locate the gaps.
We begin where Thai breaks the defaults: tokenization, data, and evaluation. Each failure exposes a broader gap shared by languages the frontier was not built around.
Measure first. Understand what is uniquely Thai.
We go deep on evaluations for Thai, then understand how current models represent it, and build solutions from there.
Share the method.
We turn findings into open recipes, so the next team facing the same challenges can start from evidence, not zero.
A large voice-cloning model as a data source: a 15-second reference becomes an 82M-parameter fixed-voice student, trained entirely on synthetic speech and running on-device.
Pull synthetic realism apart and what transfers is typeface diversity, two-dimensional structure, real handwriting glyphs — enough for Wayu-Paxa-OCR-Zero, with no real OCR labels.
Wayu is a non-profit research initiative. No one understands Thai better than Thais, and no one else is coming to do this work — so we run the experiments and turn what we learn into methods any team, in any language, can use to compete.
Contribute a problem or become a research partner.