Model Card for Vintix

This is a multi-task action model via in-context reinforcement learning

Model Details

Setting	Description
Parameters	332M
Model Sizes	Layers: 20, Heads: 16, Embedding Size: 1024
Sequence Length	8192
Training Data	MuJoCo, Meta-World, Bi-DexHands, Industrial Benchmark

Model Description

• Developed by: dunnolab
• License: Apache 2.0

Model Sources

• Repository: https://github.com/dunnolab/vintix
• Paper: https://arxiv.org/abs/2501.19400

Citation

@article{polubarov2025vintix,
  author={Andrey Polubarov and Nikita Lyubaykin and Alexander Derevyagin and Ilya Zisman and Denis Tarasov and Alexander Nikulin and Vladislav Kurenkov},
  title={Vintix: Action Model via In-Context Reinforcement Learning},
  journal={arXiv},
  volume={2501.19400},
  year={2025}
}