cs.AIFeb 19, 2026
SaveTexo: Formula Recognition within 20M Parameters
Organizations: LTCI, T´el´ecom Paris, Institut Polytechnique de Paris Paris, France
Abstract
In this paper we present Texo, a minimalist yet highperformance formula recognition model that contains only 20 million parameters. By attentive design, distillation and transfer of the vocabulary and the tokenizer, Texo achieves comparable performance to state-of-the-art models such as UniMERNet-T and PPFormulaNet-S, while reducing the model size by 80% and 65%, respectively. This enables real-time inference on consumer-grade hardware and even in-browser deployment. We also developed a web application to demonstrate the model capabilities and facilitate its usage for end users.
Figures & tables
Fig. 1: Example images that Texo is able to recognize.
| Model | GOT-OCR2.0 | UniMERNet-T | PPFormulaNet-S |
|---|---|---|---|
| Encoder | ViT 95M | SwinT 25M | HGNetV2 14M |
| Decoder | Qwen2 463M | MBart 81M | MBart 44M |
| Vocab. size | 152K | 50K | 50K |
| Hidden. dim. | 1024 | 512 | 384 |
| Embedding | 155M | 51M | 38M |
| Total | 560M | 107M | 58M |
TABLE I: Parameter counts of formula recognition models
Fig. 2: Texo model architecture in the training process
| Model | Params. | CDM ↑ | Token length ↓ | GPU inference time per sample(ms) ↓ | ||||||
| SPE | CPE | SCE | HWE | SPE | CPE | SCE | HWE | |||
| Mathpix | - | 0.973 | 0.967 | 0.924 | 0.932 | - | - | - | - | - |
| UniMERNet-T | 107M | 0.991 | 0.949 | 0.938 | 0.933 | 1 14.82 | 475.52 | 3 5.93 | 3 4.99 | 2266.96 |
| PPFormulaNet-S | 5 7M | 0.949 | 0.678 | 0.856 | 0.818 | 117.10 | 4 61.16 | 45.95 | 41.44 | 217.60 |
| Texo | 20M | 0 .958 | 0 .825 | 0 .882 | 0 .902 | 62.63 | 242.77 | 20.92 | 21.48 | 3 11.26 |
TABLE II: The recognition performance of different models on the UniMER-Test dataset.