Kyutai's REKI Hits 90.4% on LIBERO-Plus With 2.4B Parameters

Kyutai, together with Inria and École des Ponts, has released REKI, a 2.4B-parameter vision-language-action model (VLA, a model that maps camera views and an instruction directly to robot actions) that scores 90.4% on LIBERO-Plus — the top score in Kyutai's comparison, ahead of models nearly twice its size. All weights are live on Hugging Face today under a CC BY-NC-SA 4.0 license. The model was pretrained exclusively on public data.

The release matters for two reasons. First, memory: REKI carries a cheap record of its recent past, which most VLAs lack, and that memory costs 0.1 milliseconds per action chunk. Second, efficiency: the model runs a full action chunk in 21.8 ms compiled on a single NVIDIA H100, versus 33.8 ms for Physical Intelligence's π0.5 and 46.1 ms for AMAP's ABot-M0, both measured by Kyutai.

What is REKI?

REKI pairs a Qwen3.5-2B vision-language backbone with a 200M-parameter action expert, for 2.4B parameters in total. The action expert generates action chunks by flow matching (a generative method that learns to transform noise into data along a learned trajectory), integrated in 10 Euler steps. Robot state and flow time enter the model through adaptive RMSNorm, a normalization layer conditioned on those inputs. Each robot state layout gets its own small state MLP, so the same backbone serves different robot embodiments.

Kyutai benchmarked the model on one NVIDIA H100, compiled, with two cameras at 256×256 resolution. Weights ship in BF16. No quantized variants were announced.

How does REKI's memory work?

Most VLAs act on the current frame alone. Prior memory approaches feed past frames into the VLM, unroll the training sequence, or bolt on extra modules — and each option adds cost, according to Kyutai.

REKI takes a different route. It places memory inside the action expert, not the VLM. Before each frame's action tokens sit 32 learned register tokens (small learned vectors that summarize a frame). These registers cross-attend to the encoded prompt and the views of that frame. They also self-attend to the registers of earlier frames. The memory covers the last 3 frames, sampled 1 second apart.

Crucially, registers never read action tokens. They therefore depend only on observations, not on flow time or robot state, and the system computes them once per frame and caches them in a rolling window. The cost, per the project page:

• Reading cached registers adds 0.1 ms to a 21.8 ms chunk.
• The cache uses about 2 MiB of GPU memory per past frame.
• Training packs a 3-frame window into a single batched pass.

The memory also fixes a concrete failure mode. Without it, robot arms often oscillate between two poses. On RoboTwin, oscillating episodes fall from 18.6% to 4.1% with memory enabled. Kyutai attributes the oscillation to perceptual aliasing (two visually similar states requiring different actions).

How does REKI perform on benchmarks?

Kyutai pretrained REKI once, then finetuned it separately per benchmark. Scores come from the model card and the project page:

• LIBERO-Plus: 90.4%, ahead of Qwen-RobotManip (89.0%) and π0.5 (85.7%, per Kyutai's evaluation).
• RoboTwin Clean2Rand: 65.8%, behind only Qwen-RobotManip (68.5%).
• RoboCasa365: 38.4%, ahead of RLDX-1 (36.0%) and Qwen-RobotManip (35.9%), but well behind Xiaomi-Robotics-1 at 57.4%.
• LIBERO (4 suites): 97.6%.
• RoboTwin 2.0 Easy / Hard: 85.9% / 84.6%.

The ablations quantify where the gains come from. Pretraining helps most out of distribution: it adds 13.8 points on RoboTwin Clean2Rand and 30.3 points under clutter. Memory adds 1.5 points on LIBERO-Plus and 5.4 points on RoboCasa365.

On three out-of-distribution benchmarks, only models about twice REKI's size beat it, per Kyutai's comparison — Qwen-RobotManip at 4.7B, π0.5 at 3.3B and ABot-M0 at 4.6B, all sized by Kyutai.

How does REKI compare with other VLAs?

Feature	REKI (Kyutai)	Qwen-RobotManip (Qwen)	π0.5 (Physical Intelligence)	ABot-M0 (AMAP)
Total size	2.4B	4.7B	3.3B	4.6B
VLM backbone	Qwen3.5-2B	Qwen3.5-4B	Gemma-based	Not disclosed
Pretraining data	Public only	Open-source + human video	Includes own robot data	Open-source (UniACT)
Memory	32 registers, 3 frames	In-context variant	None	None
LIBERO-Plus	90.4	89.0	85.7	80.5
RoboTwin Clean2Rand (mean)	65.8	68.5	63.8	55.7
RoboCasa365	38.4	35.9	16.9	Not disclosed
H100 latency, compiled	21.8 ms	Not disclosed	33.8 ms	46.1 ms
Weights	Open, CC BY-NC-SA 4.0	Not released	Open, Apache-2.0 repo	Open, Apache-2.0 repo

Notes from the source: Clean2Rand means are computed from the 6 conditions in Table 5 of the Qwen-RobotManip report, and that report lists π0.5 at 84.4 on LIBERO-Plus, slightly below Kyutai's own 85.7 figure.

The comparison stakes are straightforward. Qwen-RobotManip does not release weights; π0.5 trains partly on proprietary robot data. REKI's claim to relevance is a fully public-data pipeline, open weights and the lowest disclosed latency in the table — with the caveat that its license bars commercial use.

What data was REKI trained on?

The pretraining mixture holds 1.0B frames plus 12.2M vision-language samples. Real-robot teleoperation provides 639.8M frames, simulation 193.9M frames and egocentric human video 170.9M frames. Named sources include DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse. Pretraining ran 200,000 steps on 32 H100 GPUs. One design choice stands out: Kyutai keeps each dataset's native control convention rather than remapping actions into a shared format, which the per-layout state MLPs accommodate.

How do you deploy REKI?

Six checkpoints are listed on Hugging Face: the pretrained base plus finetunes for LIBERO, RoboTwin (2 variants), RoboCasa365 and RoboCasa-GR1. Real-robot deployment uses the LeRobot async-inference gRPC protocol. Kyutai validated this on a UR5 arm, finetuning REKI for 3,000 steps — about one hour — on 50 demonstrations per task.

The inference code is short: load a checkpoint, call policy.reset() to clear the register cache at episode start, then call predict() with camera images, proprioceptive state and the task instruction.

Why does this release matter?

The robot-learning field has converged on VLAs as the default architecture for manipulation, but two constraints have limited small players: memory costs and closed training data. REKI addresses both. Its register-token memory adds 0.1 ms and roughly 6 MiB of GPU memory for a 3-frame window, a rounding error against the 21.8 ms chunk budget. And its pretraining mix is entirely public — 1.0B frames from DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse — which means any lab can audit, reproduce or extend the pipeline.

The weaknesses are equally concrete. REKI's 38.4% on RoboCasa365 trails Xiaomi-Robotics-1 by 19 points, and its CC BY-NC-SA 4.0 license excludes commercial deployment without separate terms.

Whether the register-token design becomes a standard component will depend on whether larger VLAs adopt it; for now, REKI sets the reference point for what a 2.4B, public-data-only model with sub-25 ms latency can do on out-of-distribution manipulation benchmarks.