NVIDIA Alpamayo is a family of open AI models for self-driving cars that look at camera video, explain their reasoning in text, and output a driving trajectory. NVIDIA calls them reasoning vision-language-action (VLA) models. There are three versions as of October 2026: Alpamayo 1 and 1.5 (10 billion parameters, runnable on one 24 GB GPU) and Alpamayo 2 Super (34 billion, released August 4, 2026, tested on one 80 GB H100). The weights are open under the OpenMDW-1.1 license, which NVIDIA says allows commercial use. They are teacher models that run in the data center; cars get smaller distilled versions on NVIDIA DRIVE AGX Thor, and NVIDIA expects the first production vehicles with them around late 2026 or early 2027.
Alpamayo 1, 1.5 and 2 Super compared
| Model | Released | Parameters | Backbone | Camera input | GPU to run it |
|---|---|---|---|---|---|
| Alpamayo 1 (first posted as Alpamayo-R1) | Weights on Hugging Face December 2025; launched at CES January 5, 2026 | 10B (about 8B backbone + 2B action expert) | Cosmos-Reason | 4 cameras plus ego-motion | One GPU with 24 GB+ VRAM |
| Alpamayo 1.5 | GTC, March 2026 | 10B | Cosmos-Reason2 8B | Variable camera count, navigation guidance, text prompts | One GPU with 24 GB+ VRAM |
| Alpamayo 2 Super | Announced May 31, 2026; weights August 4, 2026 | 34B (32B backbone + 2.3B action expert) | Cosmos 3 Super Reasoner | 360° surround view (tested with 7 cameras) | One H100 80 GB (peak about 72 GB) |
Specs are from NVIDIA’s Alpamayo model information page and the model cards for Alpamayo-R1, Alpamayo 1.5 and Alpamayo 2 Super. NVIDIA’s page now labels the 10B models “Nano”. Alpamayo 1 outputs a 6.4-second trajectory of 64 waypoints along with a text trace that explains the decision, which NVIDIA calls chain-of-causation reasoning. Software requirements are Linux, PyTorch 2.8 or later and recent Hugging Face Transformers; code is on GitHub.
Can you use Alpamayo commercially?
- Model weights: OpenMDW-1.1 across the family. NVIDIA says this allows commercial use, fine-tuning, derivative models and redistribution. Older Alpamayo 1 and 1.5 model cards said “non-commercial”; NVIDIA’s information page says that wording is superseded, so check the license file in the version you download.
- Code: Apache 2.0.
- Driving dataset (PhysicalAI-Autonomous-Vehicles, about 1,700 hours of driving from 25 countries): a separate NVIDIA AV dataset license that NVIDIA’s page describes as internal development only, no redistribution, with access expiring 12 months after first download.
- Model outputs: no license conditions, per NVIDIA.
The split matters if you plan to ship a product. Fine-tuning the weights for your own fleet is a different license question from redistributing the dataset or a model trained on it.
How Alpamayo reaches a car
Alpamayo itself is too large for an in-car computer. NVIDIA positions it as a teacher: developers use it to label tricky scenes, generate reasoning data and train smaller student models, then run those students on DRIVE AGX Thor inside the DRIVE Hyperion reference design. Testing happens in AlpaSim, an open-source closed-loop driving simulator (Apache 2.0), and AlpaGym adds reinforcement-learning training on top.
Who is using Alpamayo
- Uber: NVIDIA said on March 16, 2026 that Uber’s robotaxi fleet will run NVIDIA DRIVE AV software on Hyperion, using Alpamayo open models and the Halos safety system, starting in Los Angeles and the San Francisco Bay Area in the first half of 2027 and reaching 28 cities by 2028. That is a plan; no rides have started.
- Lucid, JLR and Berkeley DeepDrive: named at the CES 2026 launch. NVIDIA describes them as showing interest, not production deployments.
- Mercedes-Benz CLA: NVIDIA’s CEO said at CES that the CLA’s driver-assistance system uses NVIDIA’s stack including Alpamayo reasoning. Mercedes describes the CLA system as SAE Level 2, and NVIDIA’s own information page warns against reading its DRIVE partner list as Alpamayo deployments.
- BYD, Geely, Isuzu and Nissan: adopted DRIVE Hyperion for Level 4 vehicles in March 2026; the announcement doesn’t say they use Alpamayo.
Downloads are counted in the hundreds of thousands on Hugging Face. For how Alpamayo fits next to Cosmos, GR00T and Isaac in NVIDIA’s physical AI lineup, see our NVIDIA physical AI explainer. Other driving-AI suppliers working with automakers and Uber include Momenta.
NVIDIA model pages, information page and press releases checked October 24, 2026.