NVIDIA introduces 32-billion-parameter Alpamayo 2 Super model for level 4 robotaxi development
NVIDIA introduces Alpamayo 2 Super, a 32-billion-parameter open reasoning model for level 4 robotaxi development.
NVIDIA has introduced Alpamayo 2 Super, a 32-billion-parameter reasoning-based vision language action model for level 4 robotaxi development.
Table Of Content
The model extends the NVIDIA Alpamayo family of open AI models, simulation frameworks and physical AI datasets. It is designed to reason, plan and act across the driving stack, giving autonomous vehicle developers a foundation model that can support perception, decision-making, trajectory generation and safety validation.
The launch also includes NVIDIA AlpaGym, NVIDIA OmniDreams and new NVIDIA Omniverse NuRec models. Together, the tools are intended to support a development pipeline that spans real-world data capture, synthetic data generation, closed-loop training and in-vehicle deployment.
Alpamayo 2 Super expands reasoning across the driving stack
Alpamayo 2 Super scales the Alpamayo family from earlier 10-billion-parameter models to 32 billion parameters. NVIDIA said the larger model improves reasoning, 3D spatial understanding and trajectory prediction, particularly in rare and complex driving scenarios.
The model is built on NVIDIA Cosmos world foundation models and expands from front-focused camera inputs to full-surround perception across front, side and rear views. This gives the model broader situational awareness for driving decisions such as lane changes, merges and intersection crossings.
NVIDIA has also added Meta-Action output, which allows the model to predict higher-level driving decisions such as yielding, changing lanes or stopping. These outputs sit alongside trajectories and chain-of-causation traces, giving downstream planning systems more context for how decisions are made.
The company said Alpamayo 2 Super also supports reasoning auto-labelling with 2D grounding. This allows the foundation model to generate reasoning labels and compress annotation cycles from months to days.
“Alpamayo is the moment cars begin to safely reason, not just drive,” said Jensen Huang, founder and CEO of NVIDIA. “Only NVIDIA makes available open models, simulation, real-world data and agent skills so the entire global robotaxi ecosystem can develop level 4 capabilities that understand edge cases, explain decisions, earn trust and scale safely to millions of vehicles.”
Designed as a teacher model, Alpamayo 2 Super can be distilled into smaller models that run on NVIDIA DRIVE Hyperion and NVIDIA DRIVE AGX Thor inside the vehicle. NVIDIA said this allows downstream autonomous vehicle stacks to inherit higher-quality reasoning and perception from a single open release, rather than requiring each manufacturer to rebuild the underlying system from scratch.
Alpamayo was recently recognised by the COMPUTEX Best Choice Awards in the Vehicle Technology and Smart Cockpit category. NVIDIA said the Alpamayo platform has been downloaded close to 400,000 times since launch.
Alpamayo 2 Super is expected to be available this summer, with inference code on GitHub and model weights on Hugging Face.
AlpaGym brings closed-loop training to autonomous vehicle development
NVIDIA also introduced AlpaGym, an open source, high-throughput closed-loop reinforcement learning framework for autonomous vehicle training.
Unlike open-loop training, which evaluates models against recorded data and produces a single set of actions, AlpaGym runs models through continuous decision and observation cycles in NVIDIA AlpaSim. Each braking, steering and navigation choice affects the simulated environment, allowing models to learn from the consequences of their actions before road deployment.
NVIDIA said this approach exposes compounding errors and edge-case failures that static datasets may miss. Built on the AlpaSim microservice simulation stack and NVIDIA Omniverse NuRec, AlpaGym is intended to support scalable closed-loop reinforcement learning and refinement after open-loop pretraining.
The company is also releasing its CoC Auto-Labeling Pipeline as open source on GitHub. The pipeline generates decision-grounded and causally linked chain-of-causation labels from raw driving clips without human annotation, creating training data for embodied reasoning models at scale.
OmniDreams and agent skills support simulation workflows
NVIDIA OmniDreams adds a generative world model for photorealistic closed-loop autonomous vehicle scenario generation. The tool is designed to simulate rare and long-tail driving scenarios at scale, helping developers train and test models against situations that may be difficult to capture frequently through fleet data alone.
NVIDIA is also launching physical AI agent skills under NVIDIA Agent Toolkit for autonomous vehicle development. These include Neural Reconstruction skills powered by NVIDIA Omniverse NuRec libraries, NVIDIA OmniDreams skills for photorealistic scenario generation and AlpaGym skills for closed-loop reinforcement learning.
The Neural Reconstruction skill uses real-world fleet driving scenarios for simulation and generates synthetic training data at scale. NVIDIA said these agent skills are intended to guide developers and coding agents through simulation, data generation and training workflows for autonomous driving systems.
