Kyutai releases REKI, a 2.4B VLA model with built-in memory

Kyutai, with Inria and École des Ponts, released REKI, a 2.4B vision-language-action model for robot manipulation pretrained only on public data. It stores the last 3 frames as 32 cached register tokens inside the action expert, adding 0.1 ms to a 21.8 ms action chunk. REKI scores 90.4% on LIBERO-Plus but only 38.4% on RoboCasa365; weights are open on Hugging Face under CC BY-NC-SA 4.0.

• 2.4B parameters: Qwen3.5-2B VLM plus a 0.2B action expert
• Memory of 3 past frames costs 0.1 ms and about 2 MiB per frame
• Scores 90.4% on LIBERO-Plus and 65.8% on RoboTwin Clean2Rand
• Memory cuts arm pose oscillations from 18.6% to 4.1%