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 2.4B vision-language-action model for robot manipulation. It takes camera images plus an instruction and turns them into robot actions. The model is built from a 2.2B Qwen3.5-2B visual-language backbone and a 200M action expert, and Kyutai says the whole thing was pretrained only on public data.

The unusual bit is the memory. Instead of bolting on a heavy external system, REKI keeps a cheap rolling cache of the last three frames, one second apart. That memory lives inside the action expert as 32 learned register tokens, which can look back at earlier frames without reading action tokens. In Kyutai’s setup, the extra cost is tiny: 0.1 ms on top of a 21.8 ms action chunk, with about 2 MiB of GPU memory per past frame.

And it seems to matter. On RoboTwin, oscillating episodes drop from 18.6% to 4.1% when memory is enabled. Kyutai says that addresses perceptual aliasing, the annoying robot habit of wobbling between two poses because the scene looks too similar from one moment to the next. The model was pretrained on a mixture of 1.0B frames and 12.2M vision-language samples, including real-robot teleoperation, simulation, and egocentric human video.

On benchmarks, REKI does best among VLAs pretrained on public data in Kyutai’s comparison on LIBERO-Plus and RoboTwin Clean2Rand. It scores 90.4% on LIBERO-Plus, 65.8% on RoboTwin Clean2Rand, and 38.4% on RoboCasa365. That last number is far below Xiaomi-Robotics-1’s 57.4%, so this is not a clean sweep. But the interesting part is how competitive a public-data model can be at this size, and how far a small memory trick can go. All weights are live on Hugging Face, with BF16 weights and no quantized version announced.