Kyutai open-sources REKI, a 2.4B VLA trained on public data only, setting a new size-performance Pareto frontier
MikePFrank · x · 2026-10-17
Kyutai (with Inria and ENPC) introduces REKI, a 2.4B vision-language-action model pretrained exclusively on public data, claiming a new Pareto frontier between performance and model size.
- Benchmarks: tops LIBERO-Plus; on RoboTwin Clean2Rand and RoboCasa365 only models roughly twice its size outperform it, beating π0.5, GR00T-N1.5/1.7, Qwen-RobotManip and others
- Architecture: a VLM followed by a flow-matching action expert, trained on a curated mix of public robot, egocentric human, and vision-language data
- Memory: a few register tokens in the action expert summarize VLM output per frame and persist in its cache at negligible cost — improving success rates in sim and on real robots, and making pose oscillations 4x rarer
- Efficiency: fine-tuned on a UR5 arm with just 50 demos per task and 3k steps (1 hour of training)
- Fully open: code for data curation, training, inference, evaluation, deployment, plus all pretrained and finetuned weights
More from Embodied
- Boston didn't get the memo: Massachusetts remains a top robotics hub — lukas_m_ziegler · 2026-10-17
- visloc-rs: pure Rust SfM, VI-SLAM and 3D Gaussian Splatting hits 100 stars — rsasaki0109 · 2026-10-17
- Humanoid robots for elder care and fall recovery 'within 10 years,' predicts founder — realmeetjames · 2026-10-17
- micro1 hits $1B ARR selling robotics data: 'it's almost entirely a data problem' — Exp_Mark · 2026-10-17
- Claude Code turned a Nothing Phone (1) into a daily-driver mainline Linux phone — notrealdotmp4 · 2026-10-17
- Scalpers in China use Apple Vision Pro multi-window trick to snag the $2,000 iPhone Duo — AIFlow_ML · 2026-10-17