# NVIDIA releases Alpamayo 2 Super weights for autonomous-driving developers > Source: > Published: 2026-08-04 17:25:29+00:00 [NVIDIA](https://www.nvidia.com/en-us/solutions/autonomous-vehicles/alpamayo/) CEO and co-founder [Jensen Huang (@JensenHuang)](https://x.com/JensenHuang) released Alpamayo 2 Super's model weights and inference code on August 4th, giving autonomous-vehicle developers access to a 34-billion-parameter system designed to perceive road scenes, explain its decisions and generate driving trajectories. Huang described Alpamayo 2 Super [in a post on X](https://x.com/jensenhuang/status/2084656303046332747?s=46) as a reasoning backbone for robotaxis, trucks, shuttles and delivery vans. NVIDIA introduced the model at GTC Taipei on May 31st, when it said the weights and code would arrive during the summer. The August 4th release turns that earlier preview into a product developers can download and test. Huang co-founded NVIDIA in 1993 after engineering roles at AMD and LSI Logic. His latest expansion of NVIDIA's autonomy business follows the same platform strategy that pushed the chipmaker beyond graphics processors and into the software, models and developer tools running on its hardware. Alpamayo puts NVIDIA further inside the decision-making layer of autonomous vehicles, where automakers and robotaxi operators have traditionally built proprietary systems. ### A large teacher model for driving systems [Alpamayo 2 Super](https://huggingface.co/nvidia/Alpamayo2-Super) combines a 32-billion-parameter Cosmos 3 Super Reasoner backbone with a 2.3-billion-parameter diffusion-based action expert. The model processes multi-camera video, text instructions and a vehicle's recent motion history. It returns predicted trajectories alongside text outputs that can include causal reasoning traces, visual answers and high-level driving decisions such as yielding, changing lanes or stopping. The public notebook configurations use six cameras and four historical frames from each camera. NVIDIA's model card says its trajectory interface generates 64 waypoints covering the next 6.4 seconds in 0.1-second intervals. The company tested the model on a single H100 GPU with 80GB of memory, where the measured configuration consumed roughly 69.5 GiB of reserved GPU memory. That compute requirement makes Alpamayo 2 Super primarily a cloud-based development and training model rather than software intended to run unchanged inside a vehicle. NVIDIA is positioning the release as a teacher model. Developers can use its outputs to label fleet data, evaluate smaller driving models or distill its reasoning into systems that fit the latency and compute limits of an onboard computer such as DRIVE AGX Thor. The [inference code and related development recipes](https://github.com/NVlabs/alpamayo-recipes) also cover fine-tuning, reinforcement-learning post-training and quantization. The weights are distributed under the OpenMDW 1.1 license, while NVIDIA says the source code uses Apache 2.0. NVIDIA's product page describes Alpamayo 2 Super as available for commercial use, although deploying a model inside a production vehicle still requires system-level testing, safety validation and integration with the rest of the autonomy stack. ### NVIDIA moves further up the autonomy stack Alpamayo 2 Super gives NVIDIA a route into autonomous-driving programs without operating its own robotaxi network. Waymo, Tesla, Aurora, Waabi and other autonomy developers have built vertically integrated stacks around their own data, models and vehicle programs. NVIDIA is selling the common infrastructure beneath those efforts: accelerated computing, simulation, training tools, open models and in-vehicle hardware. That strategy lowers the starting cost for automakers, suppliers and startups that cannot justify building every part of an autonomy system internally. It also gives NVIDIA influence over the interfaces connecting training data, simulation, reasoning models and onboard deployment. The model supports 360-degree camera input and can produce what NVIDIA calls Chain-of-Causation traces, which connect an observed road condition to a proposed action. Those explanations can help engineers inspect failures and compare a smaller deployed policy with the larger teacher model. They do not establish that the model is safe enough for unsupervised road operation. NVIDIA reported several evaluations in its [technical release](https://developer.nvidia.com/blog/generate-trajectories-reasoning-traces-and-auto-labels-with-nvidia-alpamayo-2-super/), including a score of 79.2 on LingoQA and an AlpaSim closed-loop score of 1.50 across 913 reconstructed scenes. The results come from NVIDIA's own testing and do not provide a direct comparison of disengagements, collisions or autonomous miles against commercial driving systems. The model card says Alpamayo 2 Super was trained with roughly 115,000 hours of multi-camera driving video and about 3.7 million generated reasoning traces. NVIDIA had previously said the broader Alpamayo platform approached 400,000 downloads, a figure that combines releases across the product family and does not show how many teams have placed an Alpamayo-derived system in a vehicle. The August 4th release gives developers enough material to test NVIDIA's central bet: that autonomous-driving teams will adopt a shared reasoning foundation and reserve their proprietary work for fleet data, post-training, safety systems and vehicle integration. If that approach takes hold, NVIDIA will supply far more than the GPUs used to train autonomous vehicles. It will provide the model architecture that helps decide how they move.