REKI 2.4B
A vision-language-action (VLA) robot policy model released by Kyutai in collaboration with Inria. It uses a VLM backbone plus a Flow Matching action expert architecture, implements short-term memory through register tokens, and supports high-frequency closed-loop control.
Timeline (last 90 days)
Kyutai, together with Inria and École des Ponts, released the REKI 2.4B model, with weights now available for download on Hugging Face.
REKI 2.4B has 2.4B total parameters, consisting of a 2.2B-parameter vision-language backbone and a 200M-parameter Flow Matching action expert.
REKI uses 32 learnable register tokens to implement short-term memory inside the action expert, covering the stereo observations of the most recent 3 seconds.
The register tokens add only about 2 MiB of VRAM and 0.1 milliseconds of incremental inference latency per frame.
REKI's register tokens are constrained by an attention mask and never read action tokens, so they only need to be computed once before the Euler integration loop begins.
REKI pretraining used 1 billion frames of visual data and 12.2 million vision-language multimodal samples, training for 200,000 steps on 32 H100s.
All of REKI's pretraining data comes from public datasets, including DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse.
REKI achieves 90.4% on the LIBERO-Plus benchmark, surpassing Qwen-RobotManip (89.0%) and π0.5 (85.7%).
REKI reduces the oscillation episode rate on the RoboTwin benchmark from 18.6% to 4.1%.
On a single H100 with torch.compile, BF16 precision and two 256x256 RGB inputs, REKI produces a complete action chunk in only 21.8 milliseconds, 35.5% faster than π0.5 (33.8 ms).
4 more timeline events
All Facts (14)
REKI pretraining preserves each dataset's native control convention, and each robot embodiment enters the action expert through a dedicated state MLP plus AdaRMSNorm.
50%UnverifiedREKI 2.4B has 2.4B total parameters, consisting of a 2.2B-parameter vision-language backbone and a 200M-parameter Flow Matching action expert.
50%UnverifiedREKI uses 32 learnable register tokens to implement short-term memory inside the action expert, covering the stereo observations of the most recent 3 seconds.
50%UnverifiedThe register tokens add only about 2 MiB of VRAM and 0.1 milliseconds of incremental inference latency per frame.
50%UnverifiedREKI's register tokens are constrained by an attention mask and never read action tokens, so they only need to be computed once before the Euler integration loop begins.
50%UnverifiedREKI pretraining used 1 billion frames of visual data and 12.2 million vision-language multimodal samples, training for 200,000 steps on 32 H100s.
50%UnverifiedAll of REKI's pretraining data comes from public datasets, including DROID, BridgeData V2, AgiBot World, EgoDex and EgoVerse.
50%UnverifiedREKI achieves 90.4% on the LIBERO-Plus benchmark, surpassing Qwen-RobotManip (89.0%) and π0.5 (85.7%).
50%UnverifiedREKI reduces the oscillation episode rate on the RoboTwin benchmark from 18.6% to 4.1%.
50%UnverifiedOn a single H100 with torch.compile, BF16 precision and two 256x256 RGB inputs, REKI produces a complete action chunk in only 21.8 milliseconds, 35.5% faster than π0.5 (33.8 ms).
50%UnverifiedREKI can converge in fine-tuning on real hardware such as the UR5 with only 50 demonstrations and 1 hour of training (3000 steps).
50%UnverifiedThe single-frame memoryless policy has a RoboTwin oscillation rate of 18.6%, injecting multi-frame history into the VLM backbone gives 5.8%, and REKI gives 4.1%.
50%UnverifiedREKI uses Flow Matching to generate actions and uses 10-step Euler integration at inference time.
50%UnverifiedKyutai, together with Inria and École des Ponts, released the REKI 2.4B model, with weights now available for download on Hugging Face.
50%