Key points
- Kyutai released REKI, a 2.4B VLA model that uses 32 register tokens for efficient short-term memory.
- The model reduces robotic arm oscillation from 18.6% to 4.1% on RoboTwin benchmarks.
- REKI is pretrained on public data and achieves top scores on LIBERO-Plus.
The short version
- Kyutai, Inria, and École des Ponts released REKI, a 2.4B parameter vision-language-action model for robot manipulation.
- The model uses a novel memory mechanism with 32 register tokens to track recent visual observations without heavy compute costs.
- REKI outperforms larger models on some benchmarks while using only public data for pretraining.
The details
Kyutai, in collaboration with Inria and École des Ponts, released REKI as a vision-language-action model designed for robot manipulation tasks, according to MarkTechPost. The model contains 2.4 billion parameters, combining a 2.2B Qwen3.5-2B visual-language backbone with a 200M-parameter flow-matching action expert. Unlike many competitors that rely on proprietary datasets, REKI was pretrained exclusively on public data.
The weights are available on Hugging Face. The architecture maps camera inputs and text instructions to physical robot actions. The action expert generates movement chunks using flow matching integrated in 10 Euler steps. Robot state and flow time enter the system through adaptive RMSNorm, with specific small state MLPs handling different robot layouts.
REKI introduces a memory system within the action expert to address the limitations of processing single frames. The system uses 32 learned register tokens that cross-attend encoded prompts and current views. These registers also self-attend to registers from previous frames but never read action tokens, ensuring they depend only on observations. This design allows the registers to be computed once per frame and cached in a rolling window. Reading these cached registers adds only 0.1 ms to the 21.8 ms per action chunk latency. The cache consumes approximately 2 MiB of GPU memory for each past frame. Training packs a three-frame window into a single batched pass.
On performance benchmarks, REKI shows strong results for its size. It achieved 90.4% on LIBERO-Plus, the top score in Kyutai’s comparison, beating Qwen-RobotManip and π0.5. On RoboTwin Clean2Rand, it scored 65.8%, trailing only Qwen-RobotManip. However, it scored 38.4% on RoboCasa365, significantly behind Xiaomi-Robotics-1. The model runs on a single NVIDIA H100, processing 256x256 inputs from two cameras. Weights ship in BF16 with no quantized variants announced.
The context
Most vision-language-action models act on the current frame alone, which can lead to instability. Prior memory methods often feed past frames directly to the visual-language model, unroll training sequences, or add heavy modules, increasing computational cost. REKI’s register approach offers a cheaper alternative. This memory mechanism also addresses a specific failure mode where robotic arms oscillate between two poses. On RoboTwin benchmarks, oscillating episodes dropped from 18.6% to 4.1% when the memory was enabled. Kyutai attributes this improvement to resolving perceptual aliasing. The model is pretrained once and then finetuned separately for each benchmark.
What to watch
- Monitor for quantized variants of REKI, as only BF16 weights are currently available.
- Watch for updates on performance in more complex, unstructured environments like RoboCasa365.
- Observe how other vendors integrate similar register-token memory into their VLA architectures.
- Check for community contributions or forks of the REKI model on Hugging Face.
What it means for organisations
Teams evaluating robot manipulation systems should assess how memory handling affects operational stability. Organizations need to verify whether their hardware can support the model's specific latency and memory requirements. Suppliers should be asked about the data sources used for pretraining to ensure compliance with internal policies. Security and operations leaders must coordinate to understand the risk of deploying models that rely on public data. Internal documentation should reflect how the model handles perceptual aliasing to prevent unexpected mechanical oscillation during deployment.
What to do and how to stay safe: Kyutai
- Verify that your GPU infrastructure can handle the BF16 precision and memory requirements of the model.
- Test the model in a sandbox environment to confirm it reduces oscillation in your specific robotic setups.
- Review the public data sources used for pretraining to ensure they align with your organization's compliance standards.
- Document the model's performance limitations, particularly in complex environments, to manage operational expectations.
Step-by-step guide: Secure Home Routers for Small Business Networks
General security guidance from the TechSec Headlines newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does REKI handle memory differently from other VLA models?
REKI uses 32 learned register tokens within the action expert that cache visual observations, avoiding the cost of feeding past frames to the main model.
What benchmarks does REKI perform best on?
REKI achieves the top score on LIBERO-Plus and performs well on RoboTwin Clean2Rand, but lags on RoboCasa365.
Is REKI available for immediate use?
Yes, all weights are live on Hugging Face, pretrained on public data and finetuned per benchmark.
