Alpamayo 2 Edge

Alpamayo 2 Edge

Reads multi-camera driving footage, a navigation instruction and motion history, then outputs a driving trajectory with a written reason, and answers questions about the scene. For autonomous vehicle developers and researchers.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Alpamayo 2 Edge is a vision-language-action model from NVIDIA for autonomous driving research. It reads camera images, a text instruction and the vehicle's recent motion, then returns a planned path and a written explanation of its reasoning.

You can also ask it open questions about what a driving scene shows. It builds on the Cosmos 3 Edge backbone. NVIDIA positions it as a base for fine-tuning, and it can be deployed in cars on the DRIVE AGX Thor platform.

Who it suits: Developers and researchers who work on autonomous vehicle systems and need a compact reasoning model for planning and decision-making.

Industry | Who uses it | How they use it | Example
Automotive | Autonomous driving software teams | Fine-tune the model for their vehicle and sensors | Adapt it to a different camera layout
Research | Autonomous vehicle researchers | Study the written reasons next to each trajectory | Compare reasoning texts across sampled candidates
Embedded systems | Teams deploying in vehicles | Run a compact model on in-vehicle hardware | Target DRIVE AGX Thor

Skip it if you want a general chatbot or an image tool, since it is built around driving scenes and camera input.

Where to use it:

• Windows, macOS or Linux with Python, to download the files
• NVIDIA DRIVE AGX Thor, the in-vehicle platform the card names for deployment
• Fine-tuning workflows built on the separate NVIDIA source code repository

Use case | Who | What they get back
Trajectory prediction | Planning engineers | 64 waypoints from 0.1 through 6.4 seconds, each with a position and rotation
Causal explanations | Researchers | Written reasoning for each trajectory candidate
Scene questions | Developers | Free-form answers about what the cameras see

How to set it up:

• Python and pip to run the download commands
• Camera frames with capture timestamps, since the prompt encodes relative times
• The inference code from the NVIDIA source repository, which the card links as its code

1. Install the Hugging Face tools.
2. Download the model files.
3. Get the inference code from the NVIDIA source repository and follow it.

pip install -U huggingface_hub
hf download nvidia/Alpamayo2-Edge

How to use it: Prepare six camera views over several time steps, a navigation instruction as text and 20 past ego poses at 0.1-second steps. The card says to turn thinking off when applying the chat template, because the template turns it on by default.

enable_thinking=False

Mode | What it does
Trajectory prediction | Predicts a path from six camera views, 22 frames in all, plus the instruction and ego poses
Free-form visual question answering | Answers open questions from the same six cameras at four frames each

Example: You supply the six camera views, an instruction such as turning at the next junction, and the recent ego poses. The model returns a 64-waypoint path out to 6.4 seconds and a text explanation of why it chose that path.

For newcomers and experts: If you are new to this, think of it as a driving planner that explains itself. Start by downloading the files and reading the NVIDIA code to see how camera frames are prepared.

If you know the field, it is a VLA model with a Cosmos 3 Edge backbone and a diffusion-based action decoder. The action expert predicts acceleration and curvature per step, a unicycle model integrates them, and each trajectory candidate carries its own sampled reasoning.

Six-camera frames and ego pose history
Navigation instruction as text
Backbone writes the driving reasoning
Action expert predicts acceleration and curvature
Unicycle model builds the trajectory

Good to know:

• Weights: about 5.3 GB in BF16, according to Hugging Face
• NVIDIA frames it as a base to fine-tune and deploy, so treat it as a starting point and not a finished driving product
• Text output varies in length, and the trajectory output holds XYZ positions and 3x3 rotation matrices

Title: Parameters by part, in billions
Language model: 1.96 billion
Vision encoder: 0.41 billion
Projector: 0.08 billion
Action expert: 0.2 billion

• License: OpenMDW-1.1 for the model weights, and Apache License 2.0 for the source code
• Cost to run: none named, though in-vehicle deployment targets NVIDIA DRIVE AGX Thor hardware
• Data it reaches: camera images, text instructions and ego motion history you supply
• Not for: the source lists no exclusions

On Hugging Face, 13 November 2026

• Task: robotics
• Parameters: 2.7B
• Weights: about 5.3 GB in BF16; about 1.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 921
• Likes: 41
• License: openmdw-1.1
• Languages: en
• Based on: nvidia/Cosmos3-Edge
• Fine-tunes: 1
• Tags: safetensors, alpamayo2_edge, alpamayo, autonomous-driving, vision-language-action, pytorch, robotics

Links

• Model: https://huggingface.co/nvidia/Alpamayo2-Edge
• Files: https://huggingface.co/nvidia/Alpamayo2-Edge/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/Alpamayo2-Edge

Then follow the usage in its model card. Nothing in this kit runs by itself.