Kyutai REKI: A 2.4B Vision-Language-Action Model Bringing

Kyutai REKI: A Compact VLA Model Built for Robot Manipulation

Kyutai, in collaboration with Inria and École des Ponts, has released REKI, a 2.4B-parameter vision-language-action model (VLA) for robot manipulation. REKI is pretrained exclusively on public data and carries a lightweight memory of its recent past. Across three out-of-distribution benchmarks, only models roughly twice its size outperform it. All weights are now available on Hugging Face.

TL;DR

• Size: 2.4B parameters (a 2.2B Qwen3.5-2B VLM plus a 0.2B flow-matching action expert). Memory spans the last three frames, spaced one second apart.
• Runs on: Benchmarked on a single NVIDIA H100. Delivers 21.8 ms per action chunk when compiled, using input from two cameras at 256×256. Weights ship in BF16; no quantized variants have been announced.
• Performance: The strongest VLA pretrained on public data on LIBERO-Plus and RoboTwin Clean2Rand.
• Best: [Truncated in source content.]

Why It Matters

REKI targets a practical middle ground in embodied AI: a model small enough to run efficiently on standard hardware, yet capable of robust manipulation. Its compact memory mechanism and public-data-only pretraining make it a compelling baseline for teams building physical AI systems in 2026, a year in which open, deployable robot foundation models are increasingly expected to ship with transparent benchmarks and accessible weights.

via MarkTechPost