REKI
Kyutai, with Inria and École des Ponts, released REKI, a 2.4B vision-language-action model for robot manipulation with a cheap memory of the last 3 frames.
REKI puts memory inside the action expert using 32 learned register tokens per frame, cached in a rolling window. Reading the cache adds 0.1 ms to a 21.8 ms action chunk on one H100. On RoboTwin, memory cut oscillating episodes from 18.6% to 4.1%.
- Date
- Saturday 17 October 2026
- Lab
- Kyutai
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| LIBERO-Plus success rate | 90.4% ahead of Qwen-RobotManip 89.0% and π0.5 85.7% in Kyutai's comparison | company |
| RoboTwin Clean2Rand | 65.8% behind only Qwen-RobotManip 68.5% | company |
| RoboCasa365 | 38.4% Xiaomi-Robotics-1 scores 57.4% | company |
| Latency per action chunk | 21.8 ms compiled, 1 H100, 2 cameras at 256x256 | company |
| Parameters | 2.4B 2.2B Qwen3.5-2B VLM plus 0.2B flow-matching action expert | company |
Weights are under CC BY-NC-SA 4.0, so non-commercial. Pretrained only on public data.
Sources
This record was checked against its sources on 18 October 2026. How we check