Kyutai unveils REKI: a 2.4B vla model enhancing robot manipulation with efficient memory

Kyutai, in collaboration with Inria and École des Ponts, has introduced REKI, a 2.4B vision-language-action (VLA) model designed to improve robot manipulation. REKI is pretrained on public data and features a low-cost memory system that retains recent past information. It outperforms models of similar size on three out-of-distribution benchmarks, with only larger models surpassing it. All model weights are now available on Hugging Face. REKI integrates a Qwen3.5-2B backbone with a 200M-parameter action expert, generating action chunks through flow matching across 10 Euler steps. The model uses adaptive RMSNorm for robot state and flow time, with each state layout having its own small state MLP. Unlike most VLAs that rely solely on the current frame, REKI embeds memory within the action expert, using 32 learned register tokens that cross-attend encoded prompts and views while self-attending earlier frames. This design reduces oscillation in robotic arms, cutting failure rates on RoboTwin from 18.6% to 4.1%. Pretraining significantly boosts performance, adding 13.8 points on Clean2Rand and 30.3 under clutter. Memory further improves results, adding 1.5 points on LIBERO-Plus and 5.4 on RoboCasa365. The model is pretrained once and fine-tuned per benchmark, with six checkpoints available, including versions for LIBERO, RoboTwin, RoboCasa365, and RoboCasa-GR1. REKI is optimized for real-robot deployment using the LeRobot protocol, with fine-tuning on a UR5 arm requiring 3,000 steps over one hour.