Texo

Why we included this project

Turning a photo of a formula into working LaTeX is a narrow job, but one that anyone who keeps meeting notes in a markup editor, writes up lecture notes, or maintains a math-heavy knowledge base ends up doing constantly. Texo does it with a deliberately small model: 20 million parameters, small enough that inference runs on a consumer GPU and even in the browser. It is distilled from a larger formula-recognition model, and the repo includes the full training pipeline and datasets, so a team can retrain or fine-tune it on its own notation or handwriting instead of treating it as a frozen black box. Measured on the UniMERNet test set, the distilled weights stay close to the bigger models even at a fraction of their size. That combination of transparency and small footprint makes it a sensible default for adding math OCR to a product without committing to a heavyweight service.