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

Kyutai, working with Inria and École des Ponts, has released REKI, a 2.4 billion parameter vision-language-action model for robot manipulation. It takes camera views and an instruction and outputs robot actions. Two things set it apart: it was pretrained only on public data, and it carries a cheap memory of its last few seconds.

Most robot models of this kind act on the current frame alone. That sounds fine until two moments look identical and the arm cannot tell whether it already did the step. REKI's answer is lightweight enough to matter in practice, and the weights are on Hugging Face now, though under a non-commercial license.

We like this release, and the reason is the memory design more than any single score. Rather than feeding past frames back through the big vision-language model, REKI puts 32 learned register tokens inside its small action expert. They summarize each frame, get cached in a rolling window covering the last three frames a second apart, and are read back for about 0.1 milliseconds on top of a 21.8 millisecond action step, per Kyutai's project page. That is memory as a rounding error.

And it fixes something real. Without memory, arms often oscillate between two poses, a failure Kyutai attributes to perceptual aliasing. On RoboTwin, oscillating episodes dropped from 18.6 percent to 4.1 percent with memory on. Anyone who has watched a robot freeze into a back-and-forth loop knows that is not a cosmetic win.

The public-data point matters too. The pretraining mix pulls from open sets like DROID, BridgeData V2, AgiBot World and EgoDex, plus egocentric human video. Showing a 2.4B model can top Kyutai's comparison on LIBERO-Plus at 90.4 percent without private robot fleets lowers the bar for labs and universities that will never own one.

Now the caution. These are Kyutai's own comparisons, and REKI is finetuned separately for each benchmark. Its weakest number, 38.4 percent on RoboCasa365, trails Xiaomi-Robotics-1 at 57.4 percent by a wide margin. Simulation suites are not kitchens. Three frames of memory is short, and long multi-step chores will need more. And the CC BY-NC-SA license means startups cannot simply ship it.

Where this leads if it scales: cheap memory plus fast finetuning, Kyutai reports about an hour on 50 demonstrations per task on a UR5 arm, puts capable manipulation within reach of far more teams. That is good for research and eventually for useful household and factory robots. It also means more arms acting in shared spaces trained on thin data. I'm watching for independent real-robot results and safety testing before anyone treats benchmark wins as readiness.

Vision-language-action models pair a vision-language backbone with an action head. REKI uses a Qwen3.5-2B backbone and a 200 million parameter flow-matching action expert. Kyutai lists six checkpoints, a pretrained base plus finetunes for LIBERO, RoboTwin, RoboCasa365 and RoboCasa-GR1, and real-robot deployment uses LeRobot's async-inference protocol.