OpenAtlas

Overview

OpenAtlas is an open effort to train a large-scale model based on recent advances in linear sequence models introduced in ATLAS: Learning to Optimally Memorize the Context at Test Time (Behrouz et al., ICML 2026) — and to release the results open source for the research community.

Linear sequence models

Built on the ATLAS architecture: efficient, high-capacity memory that learns to memorize the context at test time.

Large scale

Bringing these architectural advances from research prototypes to large-scale training.

Open source

Everything released openly — so anyone can study, reproduce, and build on the results.

Team

Ali Behrouz, Farnoosh Hashemi, Daniel Cao, and Ramin Zabih.

We are actively looking for collaborators in this project. Reach out if you are interested: alibehrouz@cs.cornell.edu.

Citation

If you build on OpenAtlas, please cite:

@misc{openatlas2026,
  title  = {OpenAtlas: An Open Large-Scale Model Built on Linear Sequence Models},
  author = {Behrouz, Ali and Hashemi, Farnoosh and Cao, Daniel and Zabih, Ramin},
  year   = {2026},
  url    = {https://abehrouz.github.io/OpenAtlas/}
}

Reference: Ali Behrouz, Zeman Li, Praneeth Kacham, Majid Daliri, Yuan Deng, Peilin Zhong, Meisam Razaviyayn, and Vahab Mirrokni. “ATLAS: Learning to Optimally Memorize the Context at Test Time.” ICML 2026.

Acknowledgements

OpenAtlas is generously supported by a Slingshots grant from the Laude Institute. We are grateful for their support of open, ambitious AI research.