Xiaomi Pushes Open-Source Drive into Autonomous AI

Xiaomi announced and open-sourced Xiaomi OneVL, a unified latent-space language-visual-autonomous-driving framework that, Xiaomi says, is the industry first to integrate VLA (vision-language-action), world models, and latent-space reasoning into one architecture. Founder and CEO Lei Jun highlighted that OneVL sets new performance ceilings on multiple inference and planning benchmarks, improving both speed and accuracy versus explicit chain-of-thought and matching “answer-only” latent CoT methods. Xiaomi published the technical report on arXiv, released model weights, and open-sourced training and inference code, inviting developers and researchers to explore applications in autonomous driving and embodied intelligence. The release signals a major open-source push from Xiaomi into driving-scale multimodal models.

Xiaomi released and open-sourced Xiaomi OneVL, a unified driving model that for the first time combines VLA (visual-language-action), world models, and latent-space reasoning into a single framework. Xiaomi says OneVL leverages XLA-style inference to boost speed and accuracy, surpassing explicit chain-of-thought approaches in precision while matching the speed of “answer-only” latent CoT methods. The model reportedly achieves SOTA results on ROADWork, Impromptu and Alpamayo-R1 benchmarks and strong performance on NAVSIM, and offers both textual and visual explanations for decisions. Xiaomi published weights, training and inference code, plus a technical report and project site on GitHub and arXiv, enabling researchers and practitioners to adopt and extend the system.