This PR adds a model card for the paper [Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning](https://huggingface.co/papers/2605.21488). It includes the license, pipeline tag, relevant links to the paper and code, and the citation.
Rewrites maze-unique/eqr.pth to current EqR-compatible release checkpoint format. Keeps only model/config/EMA/metadata and removes optimizer, carry, RNG, and dataset training state. Validated with EqR eval halt_max_steps=16 and default batch: all/exact_accuracy=0.846000, all/accuracy=0.986512, W&B run gm8mjw5l.
Release stripped Sudoku checkpoint for current EqR
Rewrites sudoku-extreme/eqr.pth to current EqR-compatible release checkpoint format. Keeps only model/config/EMA/metadata and removes optimizer, carry, RNG, and dataset training state. Validated with EqR eval halt_max_steps=16 and default batch: all/exact_accuracy=0.864139, all/accuracy=0.949290, W&B run r8glsdo4.
Make Sudoku checkpoint compatible with current EqR
Rewrites sudoku-extreme/eqr.pth to use the current EqR checkpoint structure (arch eqr@EqRModel, 13 model keys, no legacy trm_langevin_full_random_z pickle refs). Validated with EqR eval halt_max_steps=16: all/exact_accuracy=0.864364, all/accuracy=0.949354, W&B run pxiz4jv6.
