Kyutai Releases REKI, a 2.4B Robotics Model That Uses Cached Memory to Reduce Repetitive Oscillations

Kyutai, Inria, and École nationale des ponts et chaussées have released REKI, a vision-language-action model with approximately 2.4 billion parameters. The project was announced on October 16, and its paper was submitted to arXiv the day before. The release includes pretrained and task-finetuned weights, as well as code for data curation, training, evaluation, and deployment on physical robots. Official release history, paper

REKI uses Qwen3.5-2B to encode camera frames and text instructions, then uses an action expert with about 200 million parameters to generate action sequences via flow matching. Its key design adds 32 learnable scratchpad tokens per frame to summarize the current observation, with their keys and values retained in a cache for later actions to read. Because these tokens do not read action tokens, they can be reused across denoising steps, and historical images do not need to be re-encoded. Training uses a window of three frames, spaced one second apart; at inference time, the memory from the previous two frames is retained. Architecture overview

The team measured 21.8 milliseconds per action chunk on a single H100, using two 256×256 image streams, ten Euler steps, and CUDA graphs. Reading historical memory adds about 0.1 milliseconds, and each past frame adds about 2 MiB of displayed memory. These figures provide a basis for understanding the cost of short-term memory, but they should not be treated as the latency of a complete robot control loop. Inference benchmark conditions

The public repository reports success rates of 90.4% on LIBERO-Plus, 65.8% on RoboTwin Clean2Rand, and 38.4% on RoboCasa365, each using its corresponding fine-tuned model. The team’s ablation study also found that, on RoboTwin Clean, the share of episodes that repeatedly oscillated between two poses fell from 18.6% without memory to 4.1% with memory. This supports the idea that short-term observations can help resolve action ambiguity, but the results are self-reported and do not establish that all robot tasks will benefit. Evaluation code and results, memory ablation

For researchers, the released data recipe and evaluation workflow can help with replication. Deployers should note that the weights are licensed under CC BY-NC-SA 4.0. The model was primarily trained on tabletop manipulation and parallel-jaw gripper demonstrations; adapting it to a new robot requires retraining the input layer for that robot’s state configuration, and the replanning interval must also match the training setup. Key questions to watch include independent replication, latency on other hardware, and the fine-tuning costs for new grippers and tasks. Model card and limitations