During CES 2026, NVIDIA officially unveiled Alpamayo, a complete ecosystem comprised of open-source AI models, simulation tools, and real-world driving data, aiming to accelerate the development of "reasoning-capable" autonomous driving technology. This system targets the long-tail scenarios that autonomous vehicles struggle to handle—rare, complex road conditions rarely seen in past data. The goal is to enable vehicles to not only see but also understand situations, reason about causes and effects, and clearly explain why they make certain driving decisions, much like humans do.
Alpamayo Open Source Ecosystem Makes its Debut, Announcing Three Core Elements
At CES , Nvidia CEO Jensen Huang revealed the complete structure of the Alpamayo family, which includes three core elements:
VLA Model with "Thinking Process"
Fully open-source, highly realistic autonomous driving simulation system
Large-scale, cross-regional real-world driving database
Huang Renxun stated that this design is intended to address the safety and various challenges of autonomous driving in the real world when facing unpredictable situations.
(Note: VLA model, short for Vision-Language-Action, is an AI model architecture that integrates seeing, hearing, and then acting.)
The biggest pain point for self-driving tours remains the long-tail scenario, which is still a safety hurdle.
Huang Renxun pointed out that autonomous driving systems must operate under extremely diverse road conditions, and the real difficulties are often not in everyday situations, but in rare but high-risk situations, such as sudden accidents, atypical traffic behaviors, or special environmental factors.
Traditional autonomous driving architectures mostly handle "perception" and "planning" separately. This approach has limited scalability when encountering unknown or new situations. Although end-to-end learning has made progress in recent years, Huang Renxun believes that in order to truly overcome the long tail problem, the system must have "causal reasoning ability" and be able to understand the relationships between events, rather than just applying existing patterns.
Alpamayo's core concept is to make the car clear step by step.
The Alpamayo family incorporates the concept of Chain-of-Thought, creating a VLA model with reasoning capabilities. This allows the autonomous driving system to progressively deduce action logic when encountering new or rare situations. The following are Alpamayo's three main capabilities:
Visual perception : understanding the road and the surrounding environment.
Language comprehension : Grasp the context and semantic flow.
Action generation : Actual driving decisions are generated.
Jensen Huang emphasized that this design not only enhances driving capabilities but also improves the interpretability of decisions, which helps build trust in the safety of autonomous driving. The overall system is based on Nvidia's Halos safety system.
The field of AI is reaching a critical turning point, with self-driving taxis being the first to benefit.
Jensen Huang went on to say that physical AI is entering a critical turning point, when machines begin to understand, reason, and act in the real world, just as ChatGPT has changed digital AI, and self-driving taxis will be one of the first applications to benefit.
He emphasized that Alpamayo enables vehicles to drive safely in complex environments and explain their rationale for decisions, which is an important foundation for achieving scalable autonomous driving.
Three pillars in one go, creating a complete open source ecosystem
Nvidia positions Alpamayo as a "teacher model," not to be deployed directly in vehicles, but as a foundation for training, fine-tuning, and distilling other vehicle-mounted models.
(Note: Distillation here refers to using Alpamayo's reasoning capabilities to mass-produce self-driving models that can run in vehicles instantly and behave at a near-professional level.)
I. Alpamayo 1: The first self-driving VLA model with a thought process chain
Alpamayo 1 boasts 10 billion parameters and takes video as input, outputting driving trajectories and complete inference processes. It also provides open-source model weights and inference code. Currently available on Hugging Face for research and development use, future versions will expand the parameter scale, inference depth, and commercial options.
(Note: Hugging Face is known as the GitHub of the AI world, its biggest highlight being its open-source model repository, integrating a large number of models and datasets.)
II. AlpaSim: A Fully Open Source Autonomous Driving Simulation Platform
AlpaSim has been released on GitHub and supports highly realistic sensor modeling, allowing users to set up traffic behavior and closed-loop testing, as well as for rapid validation and strategy optimization.
III. Physical AI Open Datasets: Large-scale real-world driving data
Physical AI Open Datasets has a database of over 1,700 hours of driving data, covering multiple geographic regions and various environmental conditions, focusing on rare and complex scenarios, and is also available for download on Hugging Face.
Huang Renxun stated that the combination of the three can form a self-reinforcing R&D cycle, accelerating the maturity of reasoning-based autonomous driving technology.
Automakers and industry leaders have expressed their support, targeting Level 4 autonomous driving.
Several automakers and research institutions have expressed interest in Alpamayo, including Lucid , JLR , Uber , and the research firm Berkeley DeepDrive . All parties agree that AI with reasoning capabilities, open simulation environments, and high-quality data will be indispensable keys to advancing Level 4 autonomous driving.
(Note: Level 1–2 is assisted driving, Level 3 is a transitional stage, and Level 4 is when you can truly start driving without a human driver.)
Further integration with other Nvidia ecosystem members will support commercial deployment.
In addition to Alpamayo, developers can also use it with other NVIDIA platforms, including Cosmos and Omniverse, and integrate the model into the NVIDIA DRIVE Hyperion architecture, along with the DRIVE AGX Thor computing platform.
NVIDIA stated that the development process can first verify performance in a simulation environment before moving to actual commercial deployment, emphasizing both safety and scalability. This process design allows autonomous driving technology to avoid starting from scratch between the "research phase" and "commercial deployment," instead enabling it to gradually move from simulation to real vehicles within the same architecture. By repeatedly verifying safety and stability in a virtual environment before entering actual road testing, risks are reduced, scalability is improved, and automakers and developers have a clearer and more controllable path when advancing commercialization.
This article, "The Alpamayo Ecosystem Debuts: Giving AI Self-Driving Vehicles Reasoning Ability and Explaining the Reasons for Decisions," first appeared on ABMedia .
