Team-PIXEL/mpixel-4s

MPIXEL

MPIXEL, introduced in Scaling Multilingual Pretraining of Pixel Language Models, is a family of multilingual pixel language models, which operate on visually rendered text. Our main models, mpixel-base (86M parameters) and mpixel-large (307M parameters), are pretrained on sixteen languages that span ten writing systems. MPIXEL uses adaptive rendering: it renders text as grapheme bigrams by default, Chinese, Japanese and Korean text one character per patch, and Arabic and Hebrew text word by word.

This model is mpixel-4s. It follows the pixel-m4 setup and is pretrained on the same four languages (English, Hindi, Chinese and Ukrainian), but its text rendering, rotary positional embeddings, data source and per-language data amount are matched to mpixel-base to remove confounding factors.

Basic usage

To use our models, please first install our mpixel package. Here is an example of loading a pretrained model and encoding input text:

from transformers import AutoModel
import mpixel.models.auto_register  # registers MPIXEL with the Auto classes
from mpixel.processing import MPIXELProcessor
processor = MPIXELProcessor.from_pretrained("Team-PIXEL/mpixel-large")
model = AutoModel.from_pretrained("Team-PIXEL/mpixel-large")
inputs = processor(["Hello world!", "The quick brown fox jumps over the lazy dog."])
outputs = model(**inputs)
print(outputs.last_hidden_state.shape)  # torch.Size([2, 513, 768])

For further information, please take a look at our repository: https://github.com/ilkerkesen/mpixel.

Citation

If you find MPIXEL beneficial for your research, please cite it,

@misc{kesen2026scalingmultilingualpretrainingpixel,
      title={Scaling Multilingual Pretraining of Pixel Language Models}, 
      author={Ilker Kesen and Jenny Kunz and Kasra Malihi and Desmond Elliott},
      year={2026},
      eprint={2610.29681},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.29681}, 
}