A 20M-parameter LaTeX OCR that fits in your browser tab

What it does

Texo reads images of mathematical formulas and spits out LaTeX source. It’s a distilled, finetuned descendant of PPFormulaNet-S and UniMERNet, shrunk down to 20 million parameters—small enough to run in a browser via ONNX, yet competitive on BLEU and edit distance with models 3–5× its size. The project ships the full training pipeline, from data normalization to Hydra-configured experiments, aimed at reproducibility on consumer GPUs (16–40 GB VRAM).

The interesting bit

The author treats this as “a comprehensive practice”—essentially a public notebook of everything learned in school and from open source. The result is unusually pedagogical: collected KaTeX commands, sorted preprocessing scripts, and explicit notes on tokenizer choices that break metric comparability. It’s research engineering as tutorial.

Key highlights

• 20M parameters vs. 57M–107M for comparable SOTA models; BLEU scores within ~1–4% of UniMERNet-T on standard test sets
• Browser demo runs fully offline after load; ONNX export included
• Training reproducible from scratch with documented minimal specs (20G CPU RAM, 16G GPU with gradient accumulation)
• AGPL-3.0 licensed, with a quirky LGPL carve-out for one specific macOS app

Caveats

• The “transfer” variant uses a custom tokenizer, making its BLEU scores not directly comparable to the distill version or baselines—sequence length differs
• Browser demo is maintained in a separate repo (Texo-web), and the author openly admits limited frontend skills; bugs there may lag

Verdict

Worth a look if you need embeddable math OCR, want to study distillation on vision-encoder-decoder architectures, or just want a well-documented training recipe to crib from. Skip if you need guaranteed production SLAs or a polished all-in-one app.