Parcae: Scaling Laws For Stable Looped Language Models

Paper | Project Page | GitHub

Parcae is a novel stable, looped architecture for language models. Unlike traditional fixed-depth architectures that scale by increasing parameter count, looped architectures increase compute (FLOPs) by sending activations through a block of layers in a loop. Parcae addresses training instabilities in prior looped models by recasting looping as a nonlinear time-variant dynamical system and constraining the spectral norm of injection parameters.

This checkpoint is the 140M parameter version of Parcae trained on the FineWeb-Edu dataset.

Installation

To use this model, install the parcae-lm package:

pip install parcae-lm

Usage

You can load the pretrained weights using the following code:

import parcae_lm

# Load a pretrained model from HuggingFace
model = parcae_lm.from_pretrained("SandyResearch/parcae-small-140m")

Citation

@misc{prairie2026parcaescalinglawsstable,
      title={Parcae: Scaling Laws For Stable Looped Language Models}, 
      author={Hayden Prairie and Zachary Novack and Taylor Berg-Kirkpatrick and Daniel Y. Fu},
      year={2026},
      eprint={2604.12946},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2604.12946}, 
}