Kyutai and Inria release REKI, a 2.4B open VLA that adds cheap memory to robot manipulation

Kyutai, working with Inria and École des Ponts, has released REKI, a 2.4B-parameter vision-language-action model for robot manipulation. MarkTechPost reported on October 17 that the weights went live on Hugging Face the same day under CC BY-NC-SA 4.0, which allows reuse but not commercial deployment.

The architecture is conventional at the top: a Qwen3.5-2B backbone reads camera views and the instruction, and a 200M-parameter action expert generates action chunks by flow matching in 10 Euler steps. Robot state and flow time enter through adaptive RMSNorm, and each state layout gets its own small state MLP.

The change is in how it remembers. Most VLAs act on the current frame only; earlier memory approaches feed past frames back into the VLM, unroll training, or bolt on extra modules, each with a cost. REKI puts memory inside the action expert instead: 32 learned register tokens sit ahead of each frame's action tokens, cross-attending to that frame's prompt and views and self-attending to the registers of earlier frames.

Because the registers never read action tokens, they depend only on observations, not on flow time or robot state. They can be computed once per frame and cached in a rolling window, which Kyutai says costs about 0.1 ms.

That memory fixes a specific failure. Without it, arms often oscillate between two poses; on RoboTwin the share of oscillating episodes falls from 18.6% to 4.1%, which Kyutai attributes to perceptual aliasing — two different situations that look identical in the current view.

REKI scores 90.4% on LIBERO-Plus, the best in Kyutai's comparison, with only models roughly twice its size ahead of it. Ablations credit memory with a 1.5-point gain on LIBERO-Plus and 5.4 points on RoboCasa365, while pretraining matters most out of distribution: 13.8 points on Clean2Rand and 30.3 under clutter.

Pretraining uses public data only: 1.0 billion frames plus 12.2 million vision-language samples, split into 639.8 million real-robot teleoperation frames, 193.9 million simulation frames and 170.9 million egocentric human video frames, drawing on DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse. Training ran 200,000 steps on 32 H100 GPUs, and Kyutai kept each dataset's native control convention rather than remapping actions.

Six checkpoints are published, covering the pretrained base plus finetunes for LIBERO, RoboTwin, RoboCasa365 and RoboCasa-GR1, and real-robot deployment uses LeRobot's asynchronous inference gRPC protocol. On a UR5 arm, finetuning for 3,000 steps, about an hour, on 50 demonstrations per task was enough, which makes REKI interesting as a research base for teams that want to build on public data, with the non-commercial license the main limit on product use.