On November 21, 2025, Xiaomi officially released and open-sourced MiMo-Embodied, the industry's first cross-domain foundation model that integrates autonomous driving and embodied intelligence. The model and weights have been simultaneously launched on Hugging Face and arXiv.

Metaverse, Sci-fi, Cyberpunk, Painting (4) Large Model

Technical Highlights  

1. Cross-domain Capabilities: A single set of parameters supports three key tasks in embodied intelligence (affordance reasoning, task planning, spatial understanding) and three core tasks in autonomous driving (environment perception, state prediction, driving planning), achieving unified modeling for indoor interaction and road decision-making.  

2. Bidirectional Synergistic Empowerment: The model has validated the knowledge transfer and collaborative effects between indoor interaction capabilities and road decision-making capabilities, providing a new perspective for cross-scenario intelligent integration.  

3. Full-chain Optimization Reliability: Adopting a multi-stage training strategy of "embodied/driving capability learning → CoT reasoning enhancement → RL fine-tuning", significantly improving the reliability of real-world deployment.

Performance Results  

In 29 core benchmark tests covering perception, decision-making, and planning, MiMo-Embodied outperforms existing open-source, closed-source, and specialized models: achieving SOTA results in 17 benchmarks in the field of embodied intelligence, setting new records in 12 benchmarks in autonomous driving, and demonstrating excellent generalization capabilities in common visual-language tasks.

Applications and Ecosystem  

Xiaomi plans to first apply MiMo-Embodied to its self-developed robot vacuum cleaner, factory AGVs, and SU7 high-level intelligent driving systems, starting from Q1 2026 via OTA updates; meanwhile, it will open up model APIs, providing a cross-domain intelligent foundation for developers in home, mobile mobility, and manufacturing fields.