Kyutai, in collaboration with Inria and École des Ponts, has released REKI, a 2.4B vision-language-action model designed for robot manipulation. Pretrained exclusively on public data, REKI pairs a Qwen backbone with an efficient action expert and memory capabilities to achieve competitive performance on robotic benchmarks. All model weights have been made publicly available on Hugging Face.
The 2.4-billion-parameter model consists of a 2.2B Qwen vision-language model and a 0.2B flow-matching action expert, utilizing a lightweight memory system that caches observation registers from recent frames. This memory architecture allows REKI to mitigate issues like perceptual aliasing and reduce robotic arm oscillation during execution. Running on an NVIDIA H100, the system computes action chunks efficiently, achieving top scores among publicly pretrained models on benchmarks such as LIBERO-Plus and RoboTwin Clean2Rand.
This release provides the robotics community with an efficient, openly available vision-language-action model that matches the performance of significantly larger systems while maintaining low computational overhead.