Kyutai has partnered with Inria and École des Ponts to release REKI, a 2.4-billion parameter vision-language-action model designed for robot manipulation. Reported by MarkTechPost, the new model was pretrained exclusively on public data and features a low-cost memory mechanism for tracking its recent past. According to the available details, REKI outperformed models roughly twice its size across three out-of-distribution benchmarks. All model weights have been made publicly available.
Understanding Vision-Language-Action Models
Vision-language-action models like REKI represent a growing area of robotics research, aiming to bridge perceptual inputs, linguistic instructions, and physical manipulation. By incorporating memory of past states into a relatively compact 2.4-billion parameter architecture, developers seek to improve how robots handle sequential tasks without requiring massive computational footprints. Pretraining on public data also addresses transparency and data sourcing concerns common in machine learning development.
Unanswered Questions and Future Outlook
Despite the release of the model weights and benchmark results, several operational details remain absent. The announcement did not specify the exact hardware requirements needed to run REKI in real-time robot hardware, nor did it detail the specific composition of the public training dataset. Furthermore, the publication did not outline commercial use terms or licensing specifics beyond the availability of the weights. Observers will need to watch for community implementations to see how REKI performs outside of the tested benchmarks.
The release marks another step toward open-weight foundational models in robotics, though practical deployment hurdles remain to be documented.
Lead image: Magda Ehlers / Pexels
