TLDR
- Nvidia has introduced Alpamayo, a new family of open source AI models designed for autonomous systems.
- Alpamayo 1 is a 10 billion-parameter model that helps vehicles reason through complex driving scenarios.
- The model allows autonomous vehicles to explain their actions and select the safest path in real time.
- Developers can fine-tune Alpamayo and access its code through the Hugging Face platform.
- Nvidia launched Cosmos to generate synthetic environments for training and testing AV applications.
Nvidia has introduced Alpamayo, a new open source AI model family built for autonomous driving and physical robotics, offering reasoning capabilities, simulation tools, and extensive datasets, all aimed at improving safety and performance in complex environments across geographies, the company announced Monday at CES 2026.
Alpamayo 1 Introduced as Core Vision-Language-Action Model
Nvidia launched Alpamayo 1 as the foundation of its new AI model family focused on real-world decision-making for autonomous systems. The model features 10 billion parameters and uses a chain-of-thought reasoning structure to process and act on sensor inputs.
It allows autonomous vehicles to evaluate complex edge cases without prior examples, including traffic light outages or unpredictable road scenarios.
“It breaks down problems into steps, reasons through every possibility, and then selects the safest path,” said Ali Kani, Nvidia’s VP of Automotive.
The model combines vision, language, and action (VLA) to simulate human-like decision-making for steering, braking, and acceleration functions. Jensen Huang, CEO of Nvidia, explained, “It tells you what action it’s going to take, and the reasons by which it came about that action.”
Developers can access Alpamayo 1 on Hugging Face and use it to build tailored driving systems or streamline video data annotation. It supports fine-tuning into smaller and faster models, helping manufacturers accelerate integration into autonomous platforms.
Cosmos Enables Synthetic Data Training for AV Applications
Cosmos, Nvidia’s generative world model, supports Alpamayo by creating realistic synthetic environments for training and testing driving systems. This tool helps simulate rare and unpredictable conditions that traditional datasets may lack, improving system readiness.
By combining real and synthetic data, developers can test how Alpamayo-based systems behave in challenging traffic scenarios at scale. “They can use Cosmos to generate synthetic data and train Alpamayo-based AV applications,” said Kani during the press event.
Cosmos aids in building auto-labeling tools, safety evaluators, and perception systems for both physical robots and autonomous vehicles. The combination of simulation and reasoning enhances the flexibility of AI development for real-world robotic deployment.
Nvidia’s integration of Cosmos allows developers to streamline validation workflows across multiple conditions without risking real-world testing failures. This creates a safer and more efficient approach to AI training cycles in robotic systems and self-driving cars.
AlpaSim and Dataset Released for Scalable Testing
Nvidia also launched AlpaSim, a simulation framework designed to test autonomous systems in near-real-world digital environments. The company made it available on GitHub to allow developers global access to recreate diverse traffic and sensor conditions.
The framework supports scalable validation of AI models under different weather, lighting, and road configurations without real-world limitations. Developers can simulate edge cases continuously while evaluating system performance before deployment.
Nvidia released an open dataset with over 1,700 hours of driving footage captured across multiple regions and scenarios. The dataset features rare driving events to train models in conditions not commonly found in standard AV training sets.
Each component of Alpamayo is aimed at improving the precision and reasoning ability of autonomous platforms from training to deployment.
The combined toolkit and resources are now publicly available for development and customization.
Nvidia stated that Alpamayo, Cosmos, and AlpaSim are all designed to help bring physical AI into real-world, safe applications. All components are now accessible through Hugging Face and GitHub for developers, researchers, and automotive companies.
