The Pulse
Kyutai Releases REKI, a 2.4B Robot-Action Model
Kyutai released REKI, a 2.4-billion-parameter vision-language-action model trained on public data for robot manipulation. Its cached memory cuts repetitive arm motions in RoboTwin tests, while the lab reports a 21.8-millisecond action-chunk
AI.info Team · Written with AI ·
REKI predicts a robot-action chunk in 21.8 milliseconds on an Nvidia H100, according to Kyutai’s compiled inference test. The 2.4-billion-parameter model, released by the Paris-based lab on October 16 with research partners Inria and École nationale des ponts et chaussées, combines camera images and a written instruction to direct a robot. Kyutai says its model is smaller than the systems it compares against on two benchmarks, while matching or exceeding their results in several tests.
Memory built into the action expert
REKI pairs a Qwen3.5-2B vision-language model with a 200-million-parameter flow-matching action expert. The first component encodes camera views and the task instruction; the expert uses that context, along with robot state, to predict movement. Rather than add a separate memory module, REKI uses learned register tokens in the action expert to retain summaries of recent observations.
The model is trained on three-frame windows, with frames taken one second apart. At inference, it caches those summaries and carries them into later action chunks. Kyutai says a three-frame memory adds about 0.1 milliseconds to reading the cached registers and roughly 2 MiB of GPU memory per past frame.
“This memory improves success in simulation and on real robots, and makes oscillations between poses about four times rarer.”
Juliette Marrie, Paul Pacaud, Adrien Ramanana Rahary, Shizhe Chen and Moritz Böhle, authors of “REKI: Registering Knowledge in VLAs”
In the lab’s RoboTwin Clean tests, oscillating episodes fall from 18.6% without memory to 4.1% with it. Kyutai attributes the repeated back-and-forth motion to perceptual aliasing: two different poses can each look like the next step in a demonstration, leaving a memoryless policy to switch directions. A short history can distinguish between them.
Results across three robot benchmarks
Kyutai evaluates REKI on LIBERO-Plus, RoboTwin Clean2Rand and RoboCasa365, finetuning the pretrained model separately for each benchmark. LIBERO-Plus perturbs 40 tasks across camera view, robot starting state, language, lighting, background, sensor noise and object layout. RoboTwin Clean2Rand tests models trained on clean scenes under randomized conditions; RoboCasa365 evaluates kitchen tasks across atomic and composite groups.
The lab reports that REKI has the highest out-of-distribution success among models pretrained on public data on LIBERO-Plus and RoboTwin Clean2Rand, despite being the smallest in its comparisons. On RoboCasa365, REKI scores 38.4, compared with 36.0 for RLDX-1 and 35.9 for Qwen-RobotManip; Kyutai identifies both of those systems as pretrained on in-house data. Its published comparison says larger models—roughly twice REKI’s size—are the ones that outperform it on RoboTwin Clean2Rand and RoboCasa365.
The 21.8-millisecond figure is a lab benchmark, not a general measure of how quickly a physical robot completes a task. Kyutai measured one action chunk from raw observations on a single H100, using LIBERO inputs and compiled model code. In that same compiled comparison, the lab lists X-VLA at 25.5 milliseconds, π0.5 at 33.8, ABot-M0 at 46.1 and InternVLA-A1.5 at 51.3.
Public-data training, released checkpoints
Kyutai says REKI was pretrained for 200,000 steps on 32 H100 GPUs using public robot-manipulation, egocentric-human and vision-language data. The human-manipulation portion includes about 350,000 episodes and 170 million frames. The lab says it keeps each dataset’s native control convention rather than remapping every source into a shared action space.
The release includes training and evaluation code, data-processing tools, pretrained weights and benchmark-specific finetuned checkpoints. Those checkpoints target several different platforms, including the Franka Panda, ALOHA bimanual robot, mobile PandaOmron and Fourier GR-1 humanoid. Kyutai also describes autonomous rollouts on a UR5 arm, finetuned for one hour on 50 demonstrations per task. The listed checkpoints are tailored to their respective benchmarks and robots, rather than presented as one already-universal policy.