Meet Kyutai REKI: The 2.4B VLA Model Bringing Efficient Memory to Robot Manipulation

Why it matters

Kyutai, working with Inria and École des Ponts, has released REKI, an open vision-language-action model for robot manipulation that is pretrained entirely on public data. The model pairs a Qwen3.5-2B backbone with a flow-matching action expert and introduces a lightweight memory mechanism: learned register tokens that summarize recent observations and are cached in a rolling window. Because these registers never read action tokens, memory adds only trivial latency while letting the robot draw on its immediate past. Kyutai says this design addresses perceptual aliasing, a failure mode where arms oscillate between poses; on RoboTwin, memory sharply reduced oscillating episodes. REKI leads Kyutai's comparison on LIBERO-Plus among models pretrained on public data, though it trails Xiaomi-Robotics-1 on RoboCasa365. Weights are available on Hugging Face under a non-commercial license, with finetuned checkpoints for several benchmarks and a documented path to real-robot deployment.