Kyutai Releases REKI Robot Manipulation Model

REKI's architecture pairs a Qwen3.5-2B backbone with a 200 million parameter action expert that generates action chunks via flow matching. The model processes camera views and instructions to produce robot actions, with robot state and flow time integrated through adaptive RMSNorm. Each robot state configuration utilizes a dedicated small state MLP.

Unlike many VLAs that process only the current frame, REKI integrates memory within its action expert. This is achieved by using 32 learned register tokens that precede action tokens, allowing them to cross-attend to the encoded prompt and frame views, and self-attend to registers from earlier frames. These registers, which depend solely on observations, are computed once per frame and cached in a rolling window, adding minimal computational cost.

The REKI model demonstrates strong performance, achieving the top score of 90.4% on the LIBERO-Plus benchmark and performing best among VLAs pretrained on public data on RoboTwin Clean2Rand. While it was outperformed by models twice its size on some out-of-distribution benchmarks, its efficiency is notable.

The model was benchmarked on an NVIDIA H100, achieving an action chunk processing time of 21.8 milliseconds from two 256x256 cameras. The weights are provided in BF16 format, with no quantized variants announced at this time. All weights for the REKI model are now available on Hugging Face.