Kyutai REKI: 2.4B VLA Model for Robot Manipulation
Kyutai, Inria, and École des Ponts release REKI, a 2.4B vision-language-action model featuring efficient recent memory for robot manipulation tasks.
Stock photo for illustration only, not from the actual event
- Kyutai collaborates with Inria and École des Ponts to launch REKI
- A 2.4-billion parameter vision-language-action model built for robots
- Trained exclusively on public data with a cost-effective recent memory system
- Outperformed by only models twice its size across three benchmark tests
The artificial intelligence and robotics landscape is witnessing a notable advancement as Kyutai, in partnership with Inria and École des Ponts, officially unveils REKI. This newly introduced vision-language-action (VLA) model packs 2.4 billion parameters and is specifically engineered to enhance robotic manipulation capabilities with refined precision and adaptability.
A core highlight of the REKI model is its training methodology, which relies entirely on publicly available data, ensuring broad accessibility and transparency. Furthermore, the architecture integrates an economical memory mechanism to track its recent past experiences, empowering robotic systems to leverage immediate historical context for smarter, more cohesive physical actions.
Stock photo for illustration only, not from the actual event
When evaluated across three out-of-distribution benchmark tests, REKI demonstrated remarkable competency. The results highlighted that only AI models boasting roughly twice its parameter size managed to surpass its performance, underscoring the exceptional architectural efficiency achieved by the development team.
Vision-Language-Action (VLA) models represent a critical frontier in robotics by bridging visual perception, natural language understanding, and physical execution. The fact that REKI competes closely with models twice its scale while maintaining a compact 2.4 billion parameter footprint signals a broader industry shift toward efficiency. Prioritizing architectural cleverness over raw scale helps lower the hardware barrier required for real-time robotic deployment.
For researchers, developers, and robotics enthusiasts eager to experiment with the technology, all model weights are currently live and fully accessible online.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment