Alpamayo2-Super Overview

Description:

Alpamayo 2 Super is a Vision-Language-Action (VLA) foundation model with a 32B VLM backbone and a 2B diffusion-based action decoder (total 34B parameters), designed to tackle multiple autonomous vehicle (AV) development tasks. Alpamayo 2 Super was developed by NVIDIA as part of the broader Alpamayo Open Platform and can be integrated into autonomous driving software in the cloud for advanced end-to-end perception, reasoning, and motion planning.
Alpamayo2-Super was developed by NVIDIA as a part of Alpamayo.

This model is ready for commercial/non-commercial use.

License/Terms of Use:

OpenMDW-1.1

Deployment Geography:

Global

Use Case:

Developers and researchers working on autonomous vehicle systems who need a foundation model for perception, planning, and decision-making tasks. Alpamayo 2 Super supports multiple AV development tasks such as trajectory prediction, visual question answering, 2D grounding, and auto-labeling. It enables enterprises to accelerate AV software development with a unified model that combines vision-language and diffusion expertise.

Release Date:

Huggingface: 07/31/2026

Reference(s):

Alpamayo Overview
Alpamayo NV Developer Forum
OpenMDW-1.1 License
NVIDIA Vulnerability Disclosure Program
Alpamayo 2 Super Source Code
Alpamayo 2 Super GitHub Issues
Alpamayo 2 Super on HuggingFace

Model Architecture:

Architecture Type: Transformer
Network Architecture: Vision-Language-Action (VLA) model with a VLM backbone and diffusion-based action decoder.
This model was developed based on Cosmos 3 Super Reasoner with a diffusion-based action decoder..
Number of model parameters: 34B (3.4*10^10)

Input:

Input Type(s): Text, Image, Video, Other: Egomotion History
Input Format(s):

• String
• Other: RGB
• Other: Egomotion History

Input Parameters: One-Dimensional (1D), Two-Dimensional (2D), Three-Dimensional (3D)
Other Properties Related to Input: The model utilizes special tokens for sequence processing, including <|im_end|> for end-of-sequence and <|endoftext|> for padding, as defined in its tokenizer configuration.

Output:

Output Type(s): Text, Other: Trajectory
Output Format:

• Text: String
• Trajectory: Vector

Output Parameters: One-Dimensional (1D), Three-Dimensional (3D)
Other Properties Related to Output: The trajectory API returns 64 waypoints spanning 0.1 through 6.4 seconds at 0.1-second intervals. Each trajectory contains ego-frame XYZ positions and 3x3 rotation matrices. Text outputs include variable-length Chain-of-Causation reasoning, meta-actions, visual question answers, grounding coordinates, and structured auto-labeling fields.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration:

Runtime Engine(s): Transformers
Supported Hardware Microarchitecture Compatibility: : NVIDIA H100 80GB
Supported Operating System(s): Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Model Version(s):

Alpamayo 2 Super 34B trained

Integrate Alpamayo2-Super using PyTorch (≥2.8), Hugging Face Transformers (≥4.57.1), and DeepSpeed (≥0.17.4). The model can be deployed in autonomous driving software in the cloud for end-to-end perception, reasoning, and motion planning. Integration should follow V-model methodology with iterative unit and system-level testing to ensure safety and compliance.

Training, Testing, and Evaluation Datasets:

Training Dataset:

Alpamayo2 Super's training data comprises a mix of Chain of Causation (CoC) reasoning traces, Cosmos-Reason Physical AI datasets, NVIDIA's internal proprietary autonomous driving data, and public driving data.

Data Modality:

• Image (multi-camera)
• Text (reasoning traces, navigation guidance)
• Other: Trajectory data (egomotion, future waypoints)

Image Training Data Size: More than 1 Billion Images
Text Training Data Size: Less than a Billion Tokens
Video Training Data Size: 10,000 to 1 Million Hours
Data Collection Method by dataset: Automatic/Sensors
Labeling Method by dataset: Hybrid: Automated/Manually-Labelled
Properties (Quantity, Dataset Descriptions, Sensor(s)): The training dataset comprises approximately 115,000 hours of multi-camera driving video with egomotion and trajectory annotations, along with 3.7 million Chain-of-Causation reasoning traces in English. It includes machine-generated data from vehicle sensors (cameras, IMUs, GPS) and synthetic reasoning traces. Sensors used are RGB cameras, inertial measurement units, and GPS.

Testing Dataset:

Link: Proprietary autonomous driving test datasets, closed-loop simulation, on-vehicle road tests.

Data Collection Method by dataset: Automatic/Sensors
Labeling Method by dataset: Hybrid: Automated/Manually-Labelled
Properties (Quantity, Dataset Descriptions, Sensor(s)): The dataset comprises roughly 115,000 hours of multi-camera driving video with corresponding egomotion and trajectory annotations. It includes roughly 3,700,000 Chain-of-Causation (CoC) reasoning traces that provide decision-grounded, causally linked explanations of driving behaviors. Content includes machine-generated data from vehicle sensors (cameras, IMUs, and GPS) and synthetic reasoning traces. CoC annotations are in English and use a structured format that links driving decisions to causal factors. Sensors include RGB cameras, inertial measurement units, and GPS.

Evaluation Dataset:

Link: Same as Testing Dataset.

Benchmark Score: LingoQA: Lingo-Judge Score=79.2, AlpaSim: AlpaSim Score=1.50 ± 0.13, PhysicalAI-AV: minADE_6=0.911m

Data Collection Method by dataset: Automatic/Sensors
Labeling Method by dataset: Hybrid: Automated/Manually-Labelled
Properties (Quantity, Dataset Descriptions, Sensor(s)): The dataset comprises roughly 115,000 hours of multi-camera driving video with corresponding egomotion and trajectory annotations. It includes roughly 3,700,000 Chain-of-Causation (CoC) reasoning traces that provide decision-grounded, causally linked explanations of driving behaviors. Content includes machine-generated data from vehicle sensors (cameras, IMUs, and GPS) and synthetic reasoning traces. CoC annotations are in English and use a structured format that links driving decisions to causal factors. Sensors include RGB cameras, inertial measurement units, and GPS.

Inference:

Acceleration Engine: PyTorch, Transformers Hardware Requirements (GPU Architecture, Model): NVIDIA H100 80GB

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please make sure you have proper rights and permissions for all input image and video content; if image or video includes people, personal health information, or intellectual property, the image or video generated will not blur or maintain proportions of image subjects included.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Get Help

Enterprise Support

Get access to knowledge base articles and support cases or submit a ticket.