MindVLA-o1 - Li Auto's next-generation autonomous driving foundation model

MindVLA-o1 is the next-generation autonomous driving basic model launched by Li Auto. It adopts a native multi-modal MoE architecture and integrates three modes: vision, language and behavior. The model achieves spatial understanding through the 3D ViT encoder, uses the implicit world model for future prediction, and outputs driving trajectories with a unified behavior generation mechanism. Combining closed-loop reinforcement learning with software and hardware co-design, MindVLA-o1 can see further, think deeper, and act more steadily, marking a key step in the evolution of autonomous driving to universal embodied intelligence.

Main functions of MindVLA-o1

• 3D space perception : MindVLA-o1 realizes three-dimensional space perception through 3D ViT encoder and feed-forward 3DGS representation, and accurately understands the static environment and dynamic objects in the scene.
• Multimodal thinking and reasoning : The model introduces a predictive hidden world model to deduce future scene evolution in the hidden space, achieving deep integration of visual understanding and language reasoning.
• Unified behavior generation : The system uses VLA-MoE architecture and parallel decoding mechanism to generate high-precision driving trajectories that comply with dynamic constraints and meet real-time requirements.
• Closed loop self-evolution : Based on the Feed-forward scene reconstruction and reinforcement learning framework, the model continues to evolve itself in the simulation environment, breaking through the limitations of real data scale.
• Efficient device-side deployment : Through the optimization of software and hardware co-design laws, the system can be efficiently deployed on the vehicle end-side chip, taking into account model accuracy and reasoning efficiency.

Technical principles of MindVLA-o1

• 3D self-supervised visual coding : The model’s vision-centered 3D ViT encoder uses LiDAR point clouds as geometric cues, introduces feed-forward 3DGS representation to model static environments and dynamic objects respectively, and implements self-supervised training through the next frame prediction task, so that the model has both semantic understanding and three-dimensional perception capabilities.
• Predictive hidden world model : In order to avoid the high computational cost of directly generating future images, the model performs efficient predictions in a compact latent space. After three-stage training, it builds the latent space representation and deduction capabilities of future scenes, achieving the unification of understanding the present, imagining the future, and logical judgment.
• Unified behavior generation : The Action Expert in the VLA-MoE architecture specializes in driving trajectory generation. It uses parallel decoding to output all trajectory points at once to meet real-time performance. It performs multiple rounds of iterative optimization through discrete diffusion to ensure that the trajectory space is continuous and consistent with dynamic constraints.
• Closed-loop reinforcement learning : Upgrade traditional stepwise optimization reconstruction to feed-forward scene reconstruction, combine with generative models to expand simulation capabilities, and rely on a unified 3DGS rendering engine and distributed training framework to achieve a low-cost and efficient reinforcement learning closed loop.
• Software and hardware co-design : Based on the Roofline model to describe hardware constraints, it evaluates nearly 2,000 architectural configurations to find the Pareto optimal solution for accuracy and latency. It is found that wider and shallower model architectures are more efficient in the device-side scenario, shortening the architecture exploration cycle from months to days.

Key information and usage requirements for MindVLA-o1

• Positioning : Li Auto’s next-generation autonomous driving basic model, a native multi-modal VLA architecture for embodied intelligence.
• Release time : On March 17, 2026, it was officially released by Zhan Kun, the person in charge of the base model, at NVIDIA GTC 2026.
• Five major technological innovations : 3D space understanding, multi-modal thinking, unified behavior generation, closed-loop reinforcement learning, software and hardware collaborative design.
• technological evolution : From end-to-end to VLA to native multi-modality, it represents the beginning of the physical AI era.
• Application extension : The same set of VLA models can control vehicles and robots. Autonomous driving is only the starting point of physical AI.
• data level : Rely on the MindData unified VLA data engine to continuously collect, clean and automatically label large-scale driving data.
• Computing power level : Need to cooperate with MindSim controllable multi-modal world model and RL Infra reinforcement learning infrastructure to support large-scale closed-loop training.
• hardware level : Deployed based on NVIDIA Drive Orin or Thor platform, it needs to meet the Pareto optimal configuration of model accuracy and inference delay.
• simulation level : Relying on the unified 3DGS rendering engine and distributed training framework to achieve low-cost and efficient reinforcement learning iterations.

Core advantages of MindVLA-o1

• Native multi-modal unified architecture : MindVLA-o1 incorporates the three modalities of vision, language, and behavior into the same framework for joint training and alignment without post-splicing, achieving higher efficiency and better generalization capabilities.
• Deep understanding of 3D space : Through the 3D ViT encoder and feed-forward 3DGS representation, the model has both semantic understanding and three-dimensional perception capabilities, breaking through the limitations of traditional BEV flat scenes and too dense OCCs.
• Efficient deduction in latent space : The predictive hidden world model “imagines” the future in a compact hidden space, avoiding the high computational cost of directly generating images, and achieving the unity of understanding the present and predicting the future.
• Real-time and accurate decision-making : The VLA-MoE architecture combines Action Expert, parallel decoding and discrete diffusion optimization to take into account trajectory generation accuracy and real-time requirements. .
• Efficient deployment on the client side : The law of software and hardware co-design shortens the architecture exploration cycle from months to days, and finds the optimal balance of accuracy and latency on automotive chips.

Comparison of similar competing products of MindVLA-o1

Contrast Dimensions	MindVLA-o1	Tesla FSD	Huawei ADS
Architectural route	Native multi-modal VLA unified architecture	End-to-end pure vision	End-to-end + multi-sensor fusion
perception scheme	Vision-based + LiDAR geometry prompts	pure visual	Multi-sensor fusion
reasoning ability	Hidden world model predicts the future	End-to-end implicit reasoning	Rules + AI hybrid
behavior generation	MoE+parallel decoding+discrete diffusion	End-to-end direct output	piecemeal decision making
Simulation training	Feed-forward reconstruction + reinforcement learning	Shadow mode + emulation	Data closed loop is the main focus
Deployment optimization	Software and hardware co-design law	Self-developed chip Dojo/HW4.0	Ascend chip optimization
Application extension	Vehicle + Robot Universal VLA	Focus on autonomous driving	Focus on autonomous driving
technical stage	Physical AI/Embodied Intelligence	AI-based end-to-end	AI-based end-to-end

Application scenarios of MindVLA-o1

• Autonomous driving : MindVLA-o1, as the next-generation basic model of autonomous driving, can handle all-scenario driving tasks such as urban roads, highways, and complex intersections, and realize full-link intelligence from perception and understanding to decision-making and planning.
• Intelligent cockpit interaction : Relying on the language understanding capability of the native multi-modal architecture, the system can understand passengers’ voice instructions and combine with visual perception to achieve natural human-machine interaction and proactive services.
• Robot control : The same set of VLA models can be extended to the robot platform to drive different forms of embodied intelligence such as robotic arms and wheeled robots to complete tasks in the physical world.
• Simulation test verification : Generate high-fidelity virtual scenes through the MindSim world model, supporting large-scale closed-loop testing and model iteration of long-tail scenarios such as extreme weather and rare accidents.
• Intelligent traffic management : Based on 3D space understanding and prediction capabilities, it can be expanded to apply to city-level smart transportation systems such as vehicle-road collaboration and traffic flow prediction. ©