Alpamayo 2 Super – NVIDIA’s Open-Source Autonomous Driving AI Reasoning Model Operational

Alpamayo 2 Super is NVIDIA’s open-source autonomous driving AI reasoning model. Built on NVIDIA Cosmos 3 Super Reasoner, the model provides 360° environmental perception, advanced driving decision-making, and automatic reasoning label generation, ranking first on the LingoQA autonomous driving reasoning benchmark. The model is released on Hugging Face under the permissive OpenMDW-1.1 license, supporting fine-tuning, derivatives, and commercial redistribution. Alpamayo 1.5 / 1 are also provided for cloud development and in-vehicle distillation deployment, targeting Robotaxis and autonomous vehicles.

360° environmental perception: Delivers high-precision recognition and understanding of the full scene surrounding the vehicle, including complex real-world road conditions.
Advanced driving decision reasoning: Performs causal reasoning based on scene context to output safe and comfortable driving paths and action decisions.
Automatic reasoning label generation: Automatically generates reasoning labels for driving scenarios, helping developers quickly build training datasets.
Multimodal autonomous driving development: Supports the fusion of multimodal inputs such as vision and text, enabling unified processing of perception and decision-making tasks.
Model distillation support: Provides Alpamayo 1.5 / 1 for cloud development and efficient distillation into production-vehicle reasoning models.

Follow WeChat and reply “开源” to join the AI open-source project discussion group.
Foundation architecture: Built on the NVIDIA Cosmos 3 Super Reasoner foundation model, inheriting its multimodal understanding and long-chain reasoning capabilities.
Reinforcement learning post-training: Uses RL Post-Training to optimize the model for autonomous driving, improving causal reasoning, risk anticipation, and path-planning capabilities.
End-to-end multitask architecture: Uses a unified architecture to process core autonomous driving modules such as perception, prediction, and planning, enabling end-to-end reasoning from scene understanding to driving decisions.
Data-driven iteration: Iteratively trains on large-scale real-world driving data and simulated environments, covering rare and complex road conditions to improve generalization.
Model distillation framework: Supports distillation from Alpamayo 2 Super into the 1.5 / 1 versions, reducing model size while maintaining performance and adapting to the real-time inference requirements of in-vehicle hardware with limited computing power.

Obtain the model weights: Visit the official Hugging Face repository to download the Alpamayo 2 Super model weights and configuration files.
Confirm the commercial license: Read and comply with the terms of the OpenMDW-1.1 open-source license to ensure that fine-tuning, derivatives, and commercial deployment are compliant.
Prepare driving data: Integrate proprietary vehicle sensor data, driving strategies, and scenario annotations, and configure the training and validation environments.
Fine-tune and distill the model: Fine-tune Alpamayo 2 Super for specific domains using proprietary data, or distill it into Alpamayo 1.5 / 1 to adapt it to in-vehicle computing resources.
Integration and deployment validation: Connect the optimized model to the vehicle perception-decision-control pipeline, and verify system stability through simulation tests and real-vehicle road tests.

Open source for commercial use: Uses the Linux Foundation OpenMDW-1.1 permissive license, supporting fine-tuning, derivatives, and commercial redistribution without additional authorization.
Leading reasoning performance: Ranks first on the LingoQA autonomous driving reasoning benchmark, significantly outperforming Qwen2.5-VL 72B and Gemini 2.5 Pro in Lingo-Judge testing.
Full-stack coverage: Provides a complete model family, from high-compute cloud inference with Alpamayo 2 Super to lightweight in-vehicle deployment with Alpamayo 1.5 / 1.
Data sovereignty retained: Developers retain full control over their proprietary data and infrastructure, while the model’s openness supports the accumulation of proprietary enterprise knowledge.

Hugging Face model repository: https://huggingface.co/nvidia/Alpamayo2-Super

Robotaxi autonomous taxis: Provides real-time scenario reasoning and decision-making support for L4 autonomous driving fleets operating in complex urban road conditions.
Autonomous trucks for line-haul logistics: Enables long-distance perception and path planning on highways, improving freight safety and efficiency.
Advanced driver assistance systems (ADAS): Distilled lightweight models can be deployed in production passenger vehicles to enhance advanced driver assistance features.
Autonomous driving simulation testing: Uses automatic reasoning label generation to quickly build large-scale simulated scenario datasets and accelerate algorithm iteration.
Driving data annotation and quality inspection: Automatically generates reasoning labels for collected driving videos, reducing manual annotation costs and improving data quality.