Drive-Reason VLA 2 Super: Open-Source Release of a Reasoning Model Built for Autonomous Vehicles

The HorizenKR team has open-sourced Drive-Reason VLA 2 Super—a 7.2-billion-parameter vision-language-action (VLA) reasoning model designed specifically for autonomous driving. This model introduces a "think before acting" mechanism, adding a reasoning phase between perception and decision-making to make driving actions interpretable and auditable. The model uses the OpenMDW-1.1 license, which allows commercial use, aiming to enhance the safety and transparency of autonomous driving technology.

Traditional end-to-end autonomous driving systems directly map pixels to control commands. While effective in specific scenarios, they have a "black box" flaw—when a decision error occurs, it is difficult to trace the cause. The launch of Drive-Reason VLA is precisely to address this issue: by inserting a reasoning phase, the system understands the scene, assesses risks, and plans paths before taking action.

Drive-Reason VLA 2 Super adopts the VLA architecture (SigLIP ViT-SO400M visual module + Qwen2.5-7B language backbone + 4-layer MLP action head) with 7.2 billion active parameters. The reasoning process is: sensor data → scene encoding → thinking (scene graph, risk assessment, planning) → action output, with a 30Hz safety gate for checks. Each cycle answers three questions: What does the world look like? What problems might occur? What is the safest next plan?

The model is optimized for NVIDIA in-vehicle platforms:

Platform	Latency	Power Consumption	Notes
NVIDIA Orin (64W mode)	48ms	55W	Single camera, batch=1
NVIDIA Thor (110W mode)	22ms	110W	Multi-camera + LiDAR, batch=4
A latency of 22ms supports 45Hz operation, which is much higher than the 10Hz control frequency, leaving sufficient time for safety checks. The repository provides TensorRT engines, ONNX tools, and a secure MCU interface to lower deployment barriers.

The model uses a defense-in-depth architecture:

1. Safety Gate Mechanism: Runs at 30Hz, intercepts dangerous commands using hard-coded rules before action execution;
2. Interpretability: Each action is accompanied by reasoning basis, forming an audit log;
3. Graceful Degradation: When reasoning fails, it degrades in steps: "reasoning → coasting → safe parking" (<50ms);
4. Privacy Protection: Runs on the vehicle, raw images do not leave the vehicle, and training data is filtered via differential privacy.

The project provides LoRA adapters to support rapid fine-tuning for downstream tasks: autonomous taxis in dense urban areas, highway truck platooning, last-mile delivery, agricultural tractor operations, etc. Training supports supervised fine-tuning (SFT) and GRPO-based reinforcement learning (RLVR, for safety reward shaping).

Drive-Reason VLA follows the OpenMDW-1.1 license (allowing commercial use), and its open strategy is conducive to global safety reviews and technological progress. The repository includes a complete toolchain (reasoning, training, deployment, evaluation) and a closed-loop simulation environment. This model demonstrates the combination of large language model reasoning capabilities and physical world interaction, providing a reference for the implementation of embodied intelligence in autonomous driving scenarios.