From the source
Kyutai introduced REKI, a 2.4B-parameter vision-language-action model (VLA) pretrained on public data.
On LIBERO-Plus, RoboTwin Clean2Rand, and RoboCasa365, it is outperformed only by models about twice its size.
The model uses a memory mechanism in its action expert that adds negligible cost and reduces oscillations between poses.
Kyutai released the full code and all pretrained and finetuned model weights.