Kyutai Releases REKI: A 2.4B VLA Model with Efficient Memory for Robot Manipulation
Kyutai partners with Inria and École des Ponts to launch REKI, a compact vision-language-action model that improves robotic performance using a low-cost memory mechanism.
Key points
- REKI is a 2.4-billion-parameter vision-language-action model pretrained exclusively on public data.
- The model incorporates a lightweight memory system that tracks the last three frames to reduce robotic oscillation.
- REKI achieves top-tier scores on LIBERO-Plus and RoboTwin Clean2Rand benchmarks, outperforming larger competitors.
Kyutai, in collaboration with Inria and École des Ponts, has released REKI, a 2.4-billion-parameter vision-language-action model designed for robot manipulation. According to MarkTechPost, the model weights are now live on Hugging Face, marking the immediate availability of this new architecture for developers and researchers in the robotics field.
The REKI model distinguishes itself by carrying a cheap memory of its recent past, a feature intended to enhance manipulation tasks. Pretrained only on public datasets, REKI pairs a Qwen3.5-2B vision-language model backbone with a 200-million-parameter action expert. This action expert generates action chunks using flow matching integrated in 10 Euler steps, allowing the system to map camera views and instructions to precise robot actions efficiently.
Technical Architecture and Performance
REKI’s memory mechanism places registers inside the action expert rather than feeding past frames directly to the vision-language model, which typically adds significant computational cost. Before each frame’s action tokens, 32 learned register tokens cross-attend to the encoded prompt and current views while self-attending to registers from earlier frames. These registers depend only on observations, not on flow time or robot state, allowing them to be computed once per frame and cached in a rolling window.
This design adds only 0.1 milliseconds to the 21.8-millisecond processing time per action chunk and uses approximately 2 MiB of GPU memory per past frame. The memory covers the last three frames, spaced one second apart. Kyutai attributes the reduction in robotic arm oscillation from 18.6% to 4.1% on RoboTwin benchmarks to this fix for perceptual aliasing.
Benchmark Results and Comparisons
On three out-of-distribution benchmarks, only models about twice the size of REKI beat its performance, according to MarkTechPost. REKI achieved a score of 90.4% on LIBERO-Plus, the highest in Kyutai’s comparison, surpassing Qwen-RobotManip at 89.0% and π0.5 at 85.7%. On RoboTwin Clean2Rand, REKI scored 65.8%, placing it behind only Qwen-RobotManip at 68.5%.
However, the model struggled on RoboCasa365, scoring 38.4%, which was well behind Xiaomi-Robotics-1’s 57.4%. REKI was benchmarked on a single NVIDIA H100 GPU, processing inputs from two cameras at 256x256 resolution. The weights ship in BF16 format, and no quantized variants have been announced as of the release.
What should you do?
Researchers and developers interested in evaluating REKI should access the live weights on Hugging Face to test its performance on their specific robotic hardware. Users should verify that their systems can handle the BF16 format and the memory requirements for the rolling cache, particularly if deploying on edge devices with limited GPU resources.
What to do and how to stay safe: Kyutai
- Download the REKI weights from Hugging Face and validate compatibility with your existing robot control infrastructure.
- Test the model on a single NVIDIA H100 or equivalent hardware to benchmark latency and memory usage before scaling.
- Monitor the model’s performance on out-of-distribution tasks to identify potential failure modes similar to those seen on RoboCasa365.
- Review the model card and project page for specific finetuning instructions, as REKI requires separate finetuning per benchmark.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the parameter size of the REKI model?
REKI is a 2.4-billion-parameter model, consisting of a 2.2B Qwen3.5-2B vision-language model and a 0.2B action expert.
How does REKI’s memory system work?
REKI uses 32 learned register tokens that cache observations from the last three frames, adding minimal latency and memory overhead.
Which benchmarks did REKI perform best on?
REKI achieved the top score on LIBERO-Plus at 90.4% and placed second on RoboTwin Clean2Rand with 65.8%.
