Kyutai Releases REKI: An Open Robot Model Built Only on Public Data That Leads LIBERO-Plus

Kyutai, the Paris nonprofit AI lab, has released REKI, a 2.4-billion-parameter vision-language-action (VLA) model for robot manipulation that was pretrained exclusively on public datasets and ships with its full pipeline, from data curation to a real-robot policy server, according to the Kyutai REKI project page. Built with Inria and École nationale des ponts et chaussées, REKI posts the highest LIBERO-Plus robustness score in Kyutai's comparison, 90.4%, and adds a short-term memory that makes a common failure, robot arms shuttling between two poses, roughly four times rarer, according to the REKI paper on arXiv.

The paper appeared on arXiv on October 15, and the weights, code and project page drew their first coverage on October 17. The benchmark figures are Kyutai's own, and the rival scores in its charts are copied from those models' published results rather than re-run under identical conditions.

The more consequential part of the release is less the leaderboard position than what REKI makes possible. Most leading VLAs change their data, architecture and training supervision at the same time, and many keep their pretraining data or weights private, so nobody outside the lab can tell which choice produced a gain. REKI is designed as a reproducible control: a deliberately plain recipe, public data only, fast enough at inference that reinforcement-learning post-training and controlled ablations become affordable for groups without frontier-lab budgets.

What Kyutai Released and Why the Public-Data Constraint Matters

REKI pairs Alibaba's Qwen3.5-2B vision-language model, which reads the camera images and the text instruction, with a 200-million-parameter "action expert" that converts that understanding into motor commands, as described on the Hugging Face model card. Kyutai published six checkpoints on Hugging Face: a pretrained base plus versions finetuned for a Franka Panda arm on LIBERO, a bimanual ALOHA setup on two RoboTwin variants, a mobile manipulator on RoboCasa365 and a Fourier GR-1 humanoid on RoboCasa-GR1.

The pretraining mixture holds about 1.0 billion frames: 639.8 million from public real-robot teleoperation sets such as DROID, BridgeData V2 and AgiBot World, 193.9 million from motion-planned simulation, and 170.9 million from egocentric human video including EgoDex and EgoVerse, plus 12.2 million vision-language samples used to preserve spatial reasoning. Rather than forcing every robot into one shared action format, Kyutai kept each dataset's native control convention and told the model through its text prompt which robot and control mode a sample used, with a separate small state network for each of 14 body layouts.

That public-only constraint is the point. The paper notes that the strongest VLAs are pretrained partly on proprietary or synthesized data, and several, including Alibaba's Qwen-RobotManip, have no public weights. Earlier fully open efforts such as OpenVLA, Octo, RDT-1B, SmolVLA and Ai2's MolmoAct2 let researchers retrain models from scratch, but models built on larger in-house data kept the lead on most benchmarks. Pretraining took 200,000 steps on 32 Nvidia H100 GPUs with a global batch of 2,880 samples, and Kyutai has not published a total GPU-hour figure for that run.

How REKI's Memory Works Inside the Action Expert

Most VLAs act on the current camera frame alone. The obvious fix, feeding several past frames into the vision-language model, makes cost grow with the window because every image must be re-encoded; recurrent states must be unrolled step by step during training; and separate memory modules add components to train and serve, as the REKI paper explains.

REKI puts the memory somewhere cheaper. Before each frame's action tokens, the action expert inserts 32 learned "register" tokens, an idea borrowed from Meta and Inria researchers who added spare tokens to vision transformers to soak up global information. Each frame's registers read that frame's encoded images and instruction, and they also read the registers left behind by earlier frames. They never read the action tokens.

That one-way design is what makes the memory cheap. REKI generates actions with flow matching: it starts from random noise and refines a chunk of future actions over 10 integration steps. Because the registers ignore both the noisy actions and the refinement timestep, they need computing only once per frame, in the first step, and their attention keys and values are then reused in the remaining nine steps and stored in a rolling cache. The window covers three frames taken one second apart, meaning the current frame plus two preceding ones, and the oldest frame's registers are dropped when a new one arrives. Since cached registers were themselves computed from earlier registers, some information can survive past the window.

Alibaba's Qwen-RobotManip already places register tokens in its action expert, according to Kyutai, but discards them after each forward pass; REKI's change is keeping them. Training stays inexpensive because a three-frame window is packed into one batched pass, so it uses no more vision-model compute than three independent frames, and each frame's loss averages eight noise samples that share one encoder pass.

Read more: What actually drives robot VLA inference cost

Perceptual Aliasing Explains the Back-and-Forth Arm Failure

Kyutai's most useful finding is diagnostic. Trained without memory, REKI got stuck oscillating, with an arm reversing direction at three or more consecutive action chunks, in 18.6% of RoboTwin clean-scene episodes, including half of those that failed. With the three-frame memory, that share fell to 4.1%, according to the paper's ablation. Fewer than 5% of oscillating episodes ended in success, so the failure directly costs task completion.

The authors attribute the loop to perceptual aliasing, a problem described in reinforcement-learning research since the early 1990s: two different situations look the same to a policy that sees only the present. If pose A resembles demonstrations heading toward pose B, and pose B resembles demonstrations heading back to A, a memoryless policy alternates forever. One second of history breaks the tie because the robot can see where it just came from. In two-arm tasks the typical cycle was an arm swap, one reaching while the other retracted, which appeared in 65 of 1,000 episodes without memory and one with it.

Memory is not a cure-all. Where the memoryless model oscillated, the memory-equipped model succeeded on only 45% of the same seeds. Most of the benchmark gain also came from a single past frame, and a model trained with memory but run without it performed like a memoryless one, showing the benefit comes from reading the past at inference.

LIBERO-Plus Results Lead, but Rival Scores Were Not Re-Run

LIBERO-Plus, developed by Fudan University researchers and presented at CVPR 2026, perturbs the 40 tasks of the original LIBERO benchmark along seven axes, including camera viewpoint, initial robot pose, lighting, background, sensor noise, object layout and instruction wording, across more than 10,000 task variants. Its authors found that models scoring above 95% on standard LIBERO could fall below 30% under modest perturbations, and that many largely ignored language instructions.

Finetuned on standard LIBERO and tested on LIBERO-Plus, REKI scored 90.4%, ahead of Qwen-RobotManip at 89.0%, Physical Intelligence's π0.5 at 85.7% and InternVLA-A1.5 at 84.8%. On RoboTwin Clean2Rand, which trains on clean scenes and tests under randomized backgrounds, lighting, clutter and table height, it reached 65.8%, behind only Qwen-RobotManip's 68.5%. Memory added 1.5 points on LIBERO-Plus, 5.5 on RoboTwin clean scenes and 5.4 on RoboCasa365.

Three caveats temper those numbers. First, a paper footnote states that baseline scores were taken from published results, so differences in finetuning budgets and evaluation seeds are not controlled. Second, robustness depends heavily on training details: disabling camera dropout or image augmentation cut REKI's camera-viewpoint score by more than 20 points. Third, REKI is weaker in distribution. It scores about 97.5% on standard LIBERO, where most models sit between 97.6% and 99.4%, and 85.9% and 84.6% on RoboTwin's easy and hard settings, against 93.3% and 93.0% for the similarly sized InternVLA-A1.5. On RoboCasa365 kitchen tasks it reached 38.4%, ahead of RLDX-1 and Qwen-RobotManip but well behind Xiaomi-Robotics-1 at 57.4%, and only 14.1% on composite tasks unseen in training.

The 21.8 ms Figure Is Per Action Chunk, and Compilation Does Much of the Work

Kyutai reports that REKI predicts an action chunk, a short sequence of future motor commands, in 21.8 milliseconds on one H100 from two 256×256 camera images, faster than X-VLA at 25.5 ms, π0.5 through Physical Intelligence's openpi code at 33.8 ms, ABot-M0 at 46.1 ms and InternVLA-A1.5 at 51.3 ms. The figure is per chunk, not per individual action, and much of the speed comes from engineering rather than architecture. Without compilation, REKI, π0.5 and ABot-M0 all took 118 to 139 ms per chunk, because the GPU sat idle more than half the time waiting on the host. Compiling the whole model, memory included, into CUDA graphs made models two to seven times faster, and REKI's pure GPU time of 16.5 ms is roughly equal to the far smaller X-VLA's 16.1 ms.

Architecture still matters at the margins. InternVLA-A1.5 uses the same vision-language backbone, but its 24-layer, 0.5-billion-parameter action expert spends 34.9 ms on the 10 refinement steps, against 8.2 ms for REKI's 10-layer expert. The memory itself is close to free: reading cached registers from past frames adds 0.1 ms, the cache costs about 2 MiB of GPU memory per frame on a 4.7 GiB policy, and a separate measurement in the appendix puts the full register overhead, compared with a model that has no registers, at about 0.7 ms.

The gap with world-action models, which generate predicted video before choosing actions, is far larger. ABot-M0.5, one such model in Kyutai's tests, took 1.4 seconds per chunk even after Kyutai compiled it, roughly 60 times slower than REKI.

Read more: Open-source VLA inference reaches 38Hz on Jetson Thor

How REKI Compares With π0.5 and Other Open VLA Models

REKI's closest peers combine a vision-language model with an action expert. π0.5, at 3.3 billion parameters, is the usual open-weight reference point in that class, but Physical Intelligence pretrained it partly on its own robot data. Nvidia's GR00T N1.x models and InternVLA-A1.5 occupy similar territory. Above them sit larger systems that Kyutai includes as a gauge of the upper bound: Qwen-RobotManip at 4.7 billion parameters, which Alibaba describes as trained on more than 38,000 hours of curated open-source and synthesized human-to-robot data and whose weights are not public, plus RLDX-1 and Xiaomi-Robotics-1.

Several of the strongest leaderboard entries now take a different route entirely, building world-action models on video generators, and Kyutai concedes these models set the current state of the art on some benchmarks. REKI's argument is that their pretraining is hard to reproduce and costly to run, which makes it difficult to separate the effects of data scale, model size and world-model supervision. Kyutai says it expects REKI to gain from more data, whether real, synthetic or generated by a world model, but presents it as the plain reference those additions should be measured against.

What Developers Can Do With REKI Under a Non-Commercial License

The weights carry a Creative Commons BY-NC-SA 4.0 license, which bars commercial use and requires derivatives to be shared under the same terms, while the training and inference code is MIT-licensed. A company wanting REKI in a product would need to retrain from the public mixture using the open code, subject to each source dataset's own license terms. Kyutai offers no hosted API; the model runs on the user's own hardware, and finetuned policies connect to robots over Hugging Face's LeRobot asynchronous-inference protocol.

Adaptation is cheap by robotics standards. On a UR5 arm running a 15 Hz control loop with one fixed camera, Kyutai finetuned REKI for 3,000 steps, about an hour or 8 H100 GPU-hours, on 50 teleoperated demonstrations per task. Across five pick, place, stack and sweep tasks, REKI beat a finetuned π0.5 by 7 to 37.5 percentage points, and memory added 12 to 17.5 points on the three longer tasks while leaving the two shorter ones unchanged. Each task was evaluated on only 10 episodes, so those margins are indicative rather than precise. Benchmark finetunes take 40 to 400 GPU-hours.

Speed also shapes training. VLAs are increasingly refined with reinforcement learning in simulators or learned world models, where the policy is queried for every chunk of every rollout, so per-chunk latency multiplies across an entire training run. The model card lists practical limits: a new robot body needs a new state input layer trained from scratch, the replanning interval should match the one-second spacing of training frames, and the data is dominated by tabletop manipulation with parallel-jaw grippers.

Short Memory Works on Today's Benchmarks; Long-Horizon Memory Is Untested

REKI shifts the reference point for open robot learning. A public-data model at 2.4 billion parameters now sits at or near the top of the main robustness benchmarks in its developer's comparison, which suggests that some of the out-of-distribution lead held by larger in-house-data models comes from data and training choices rather than scale alone. It also shows that a few seconds of memory is enough to remove a specific, measurable failure mode, while the authors acknowledge that whether registers can carry much longer context, such as a task step completed minutes earlier, remains untested.

The next evidence that matters is external: independent reproductions of the LIBERO-Plus and RoboTwin numbers under matched finetuning, reinforcement-learning post-training results that start from REKI's checkpoints, and any license change that would let companies deploy the weights directly. If outside labs confirm the results, REKI could become the default baseline that new VLA designs are expected to beat.