---
license: mit
library_name: pytorch
tags:
  - jepa
  - world-models
  - model-based-rl
  - planning
  - arxiv:2607.04978
---

# Qantara headline checkpoints

Released checkpoints for **Qantara**, a compact (~21M-param) JEPA world model for
goal-conditioned planning from pixels and actions, and for our **LeWM** reproduction
baseline. These reproduce the LeWM-suite headline numbers (Table 1) from the ICML 2026
workshop paper.

- **Project page:** https://corl-team.github.io/qantara
- **Code:** https://github.com/corl-team/qantara
- **Paper:** https://arxiv.org/abs/2607.04978
- **Thread:** https://x.com/rusrakhimov/status/2074847486288806306

## Files

24 checkpoints = 2 methods × 4 environments × 3 training seeds.

| Pattern | Method |
|---|---|
| `qantara-<env>-s<seed>.ckpt` | Qantara (γ=1, λ_z=3, nulldrop=0) headline |
| `lewm-<env>-s<seed>.ckpt`    | LeWM reproduction baseline |

`env ∈ {pusht, tworoom, cube, reacher}`, `seed ∈ {11, 22, 33}`.

## Loading

Each file is a full model object (`torch.save` of a `jepa.JEPA`). Clone the code repo,
then:

```python
import torch
model = torch.load("qantara-pusht-s11.ckpt", map_location="cpu", weights_only=False)
```

The repository's `eval.py` consumes these directly. See the repo README for the full
train → eval → figure pipeline.

## Citation

```bibtex
@inproceedings{qantara2026,
  title     = {Qantara: Bridge-Flow Training for Multi-Paradigm JEPA Control},
  author    = {Rakhimov, Ruslan and Bredis, George and Maksyuta, Yuriy and Gavrilov, Daniil},
  booktitle = {ICML 2026 Workshop on Decision-Making from Offline Datasets to Online Adaptation: Black-Box Optimization to Reinforcement Learning},
  year      = {2026},
  url       = {https://arxiv.org/abs/2607.04978},
}
```