Kyutai REKI VLA: 2.4B Params Use 32 Register Tokens for Robot Memory

Kyutai, alongside Inria and École des Ponts, released REKI, a 2.4 billion parameter vision-language-action model designed for robot manipulation. The system maps camera inputs and text instructions to physical actions. According to MarkTechPost, the model is pretrained exclusively on public data and focuses on efficient memory usage to improve performance on complex tasks.

Root cause

Most vision-language-action models process only the current camera frame, ignoring recent history. This limitation can cause robots to oscillate between poses due to perceptual aliasing. REKI addresses this by embedding memory directly within its action expert module. It uses 32 learned register tokens that cross-attend to encoded prompts and current views while self-attending to registers from earlier frames.

Attack path

The memory mechanism operates by caching observations rather than actions. Registers never read action tokens, ensuring they depend solely on visual observations. They are computed once per frame and stored in a rolling window. This design allows the model to maintain context over the last three frames, spaced one second apart, without adding significant computational overhead to the action generation process.

Affected versions

The released model consists of a 2.2 billion parameter Qwen3.5-2B vision-language model backbone and a 200 million parameter flow-matching action expert. All weights are live on Hugging Face. The model ships in BF16 precision. No quantized variants have been announced. It was benchmarked on a single NVIDIA H100 GPU, processing 256x256 pixel inputs from two cameras.

Mitigation

• Deploy the model using the provided BF16 weights on compatible hardware.
• Monitor for oscillating behaviour in robotic arms during initial testing.
• Verify that the rolling window cache does not exceed GPU memory limits.
• Use the separate finetuned versions provided for specific benchmark environments.

Why it matters

Teams evaluating robotic AI systems should understand how memory mechanisms affect stability. Traditional models often lack context, leading to repetitive errors. By caching recent visual states, systems can distinguish between similar-looking scenes that require different actions. Security operations should assess whether cached data exposes sensitive environment details. Engineers must verify that memory modules do not introduce latency that disrupts real-time control loops or compromise safety protocols during physical manipulation tasks.

What to do and how to stay safe: Kyutai

• Review the model card for specific finetuning requirements per benchmark.
• Test the rolling window cache impact on your GPU memory allocation.
• Monitor action chunk latency to ensure it meets real-time control needs.
• Verify that the model handles out-of-distribution scenarios as documented.

Step-by-step guide: Password Policies Examples That Reduce Password Fatigue

General security guidance from the Vector Bulletin newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How does REKI compare to larger models?

On three out-of-distribution benchmarks, only models about twice its size beat REKI.

What is the latency of the memory module?

Reading cached registers adds 0.1 ms to the 21.8 ms per action chunk.

Does the model reduce oscillating errors?

Yes, on RoboTwin, oscillating episodes fell from 18.6% to 4.1% with memory.