Kyutai REKI: A 2.4B VLA Model Brings Efficient Memory to Robotic Manipulation

Kyutai, together with Inria and École des Ponts, has released REKI—a 2.4-billion-parameter vision-language-action (VLA) model purpose-built for robotic manipulation tasks. Its main selling points aren't parameter scale, but two things: pretraining on public data only, and an extremely low-cost "recent memory" mechanism.

A "small but sufficient" VLA model

REKI's 2.4B parameters put it in the lower-middle range of the current VLA landscape. By comparison, many similar efforts easily reach billions or more. The Kyutai team's choice is clear: no parameter arms race—instead, spend the budget on data efficiency and memory structure.

According to MarkTechPost's coverage, REKI was pretrained entirely on public data. This deserves separate mention—high-quality demonstration data in robotic manipulation tends to be held by a handful of labs and companies, so being able to train a competitive VLA on public data alone significantly lowers the barrier to reproduction and reduces compliance risk.

A VLA (Vision-Language-Action) model is a class of model that unifies visual perception, natural language instructions, and robot action outputs into a single neural network. Unlike models that only do image classification or text generation, a VLA must map multimodal inputs to executable low-level control signals (such as joint positions or end-effector displacements), and therefore typically requires behavior-cloning-style training on robot demonstration data. Its core challenges lie in cross-modal alignment, action temporal modeling, and generalization to unseen scenarios. In recent years the field has gradually shifted from early specialized policy networks to a paradigm of large-scale pretraining plus fine-tuning, but the scarcity of high-quality robot data remains a persistent bottleneck.

Low-cost memory: letting the model remember "what just happened"

REKI's most talked-about design is the "cheap memory of its recent past" it carries. Robotic manipulation is a classic temporal task: grasping, moving, placing—each step depends on the states of the previous steps. Policies that take only the current frame as input easily lose context in long-horizon tasks.

REKI's approach is to embed lightweight memory of recent history into the model, rather than bolting on a massive video-understanding module. This design orientation contrasts with "swap in a bigger model for stronger capability"—trading structural ingenuity for temporal reasoning ability while keeping inference cost within an acceptable range.

Out-of-Distribution (OOD) in robot learning refers to situations where the scene layout, object categories, lighting conditions, or instruction combinations at test time differ systematically from the training distribution. Because of the long-tail nature of the real world, robots in deployment will almost inevitably encounter inputs outside the training distribution, so OOD generalization reflects practical value more than in-distribution accuracy. Evaluation typically places the model in unseen environments or task variants and observes how much task success rate drops. The smaller the generalization gap, the closer the representations the model has learned are to the essence of the task rather than surface-level statistical correlations.

Out-of-distribution benchmarks: only models twice its size can beat it

The key measure of a VLA's generalization ability is out-of-distribution performance. According to the original announcement, on three out-of-distribution benchmarks, only models with roughly twice REKI's parameter count were able to surpass it.

This result has two implications. First, REKI has decent transfer ability on unseen scenes, objects, or instruction combinations, which corroborates its background of pretraining on public data only. Second, parameter efficiency is its real advantage—reaching a comparable level with half the size is quite friendly to the compute budgets of real deployments.

It should be noted that the original text does not list the specific names of these three benchmarks or the comparison model lineup one by one, so this conclusion is better treated as directional reference rather than a strict ranking basis.

"Embodiment" refers to the specific physical form and sensor-actuator configuration of a robot body, such as robotic arms with different degrees of freedom, grippers versus dexterous hands, mobile bases, or humanoid structures. Cross-embodiment transfer requires policies to share knowledge across bodies with different action-space dimensions, kinematic constraints, and control frequencies, and is a key obstacle to taking VLA from the lab to general-purpose robots. Common approaches use unified action representations or embodiment-specific adaptation layers, but a widely accepted evaluation standard is still lacking—something to keep in mind when gauging the practical applicability boundaries of models like REKI.

What open weights mean

All model weights are publicly available. For researchers and engineering teams in robot learning, this means they can directly fine-tune, evaluate, or build on REKI without having to train a VLA from scratch.

Combined with the "pretrained on public data only" point, REKI's openness is relatively rare in this field. It lowers the barrier to entry for VLA research and gives the community an auditable, reproducible baseline.

Some sober judgment

REKI's positioning is clear: it's not chasing first place on SOTA leaderboards, but demonstrating that under limited parameters and public data, a VLA can still have usable manipulation capability and generalization. The introduction of the memory mechanism is the most informative part—it points to a technical path different from "stacking scale."

However, the real difficulties of robotic manipulation often lie in the robustness of real hardware, error accumulation over long-horizon tasks, and transfer across embodiments. There is still a gap between benchmark numbers and real deployment, and REKI's actual performance will need more third-party reproduction to verify.

One-line Summary

Kyutai and collaborators release REKI, a 2.4B-parameter VLA model pretrained on public data only with built-in low-cost recent memory, fully open weights.

Paragraph Summary

Kyutai, together with Inria and École des Ponts, has released REKI, a 2.4-billion-parameter vision-language-action model for robotic manipulation. Its highlights aren't scale but two things: pretraining entirely on public data, significantly lowering the barrier to reproduction and data compliance risk; and a built-in lightweight memory mechanism for recent history, trading structural design rather than parameter count for temporal reasoning ability. On three out-of-distribution benchmarks, only models roughly twice its size outperform it, reflecting high parameter efficiency. All weights are public, making it a fine-tuning baseline for research and engineering teams. The article also cautions that a gap remains between benchmark numbers and real hardware deployment, and conclusions need third-party reproduction.