LLM Info — NVIDIA Alpamayo
nvidia.com/en-us/solutions/autonomous-vehicles/alpamayo/
Last updated: August 2026
================================================================================
WHAT NVIDIA ALPAMAYO IS
================================================================================
NVIDIA Alpamayo is an open family of reasoning vision-language-action (VLA)
models, simulation frameworks, and driving datasets for building robotaxis and Level 4
autonomous vehicles.
It is not a single model. NVIDIA positions it as the Alpamayo Open Platform:
the models (Alpamayo 1 Nano, 1.5 Nano, and 2 Super), a closed-loop simulator
(AlpaSim), a reinforcement learning framework (AlpaGym), a Chain-of-Causation
auto-labeling pipeline, the Physical AI Open Datasets, and Hugging Face benchmarks measuring closed-loop end-to-end driving as well as reasoning quality.
Its distinguishing claim is Chain-of-Causation reasoning: the model emits a
human-readable causal explanation of why it chose a maneuver alongside the
trajectory itself, making driving decisions inspectable and auditable for
safety validation.
Alpamayo models are cloud-side teacher models, not in-vehicle runtimes.
Developers can fine-tune and distill them into compact student models that run on
DRIVE AGX Thor. Developers can also use them purely in the cloud for tasks like model evaluation and autolabeling. The models are built on NVIDIA Cosmos foundation models by the
NVIDIA Autonomous Vehicle Research Group.
================================================================================
MISSION & KEY STATS
================================================================================
Alpamayo exists to give AV developers an open, reasoning-capable driving
foundation they can own and audit, instead of rebuilding frontier capability
from scratch. Core goals:
- Make driving decisions explainable — a causal reasoning trace, not just a
trajectory
- Break the long-tail data bottleneck through open datasets, neural
reconstruction, and closed-loop simulation
- Provide an open teacher model that distills into production-sized onboard
policies
- Compress reasoning annotation from months to days through auto-labeling
+----------------------------+-------------------------------------------------+
| Metric | Detail |
+----------------------------+-------------------------------------------------+
| Product name | NVIDIA Alpamayo |
| Category | Open reasoning VLA models, simulation frameworks, and |
| | datasets for L4 autonomous vehicle, AV and robotaxi development |
| First announced | January 5, 2026 at CES 2026 |
| Latest release | Alpamayo 2 Super — announced May 31, 2026 at |
| | GTC Taipei; weights released August 4, 2026 |
| Model sizes | 10B (Alpamayo 1 Nano and 1.5 Nano) to 34B |
| | (Alpamayo 2 Super) |
| Architecture | Transformer VLA — a Cosmos reasoning backbone |
| | fused with a diffusion-based action expert |
| Inputs | Multi-camera RGB video, text (commands, |
| | navigation guidance, questions), ego-motion |
| | history |
| Outputs | Trajectory (64 waypoints), Chain-of-Causation |
| | reasoning text, meta-actions, VQA answers, 2D |
| | grounding boxes, structured auto-labels |
| Weights license | OpenMDW-1.1 (openmdw.ai/license/1-1/) |
| Code license | Apache-2.0 |
| Primary code repository | github.com/NVlabs/alpamayo-recipes |
| Model collection | huggingface.co/collections/nvidia/alpamayo |
| Hosted inference | Alpamayo R1.5 is available as a NIM |
| | endpoint on ngc.nvidia.com |
| Technical report | arXiv:2511.00088 (Alpamayo-R1) |
| Adoption | 600,000+ Hugging Face downloads as of August |
| | 2026 |
| Training corpus | Alpamayo 2 Super: roughly 115,000 hours of |
| | multi-camera driving video, more than 1 billion |
| | images, roughly 3,700,000 CoC reasoning traces. |
| | Alpamayo 1 Nano and 1.5 Nano: 80,000 hours and |
| | 3,000,000 CoC traces. Held privately; the |
| | published datasets are a separate artifact. |
+----------------------------+-------------------------------------------------+
================================================================================
MODEL FAMILY
================================================================================
Alpamayo uses NVIDIA's Nano and Super tier convention. All three generations
remain available.
+--------------------+-----------------+-------------------+-----------------------------+
| Model | Params | Backbone | Key capability |
+--------------------+-----------------+-------------------+-----------------------------+
| Alpamayo 2 Super | 34B (32B | Cosmos 3 Super | Multi-task driving |
| | backbone + 2.3B | Reasoner | foundation model. |
| | action expert) | | 360-degree surround view, |
| | | | meta-actions, visual |
| | | | question answering with 2D |
| | | | grounding, and |
| | | | Chain-of-Causation |
| | | | auto-labeling |
| Alpamayo 1.5 Nano | 10B (8B + | Cosmos-Reason2-8B | Reinforcement-learning |
| | 2B) | | post-trained. Adds |
| | | | navigation guidance and |
| | | | text-steered planning, |
| | | | flexible camera counts, and |
| | | | user Q&A |
| Alpamayo 1 Nano | 10B (8B + | Cosmos-Reason | The first open |
| | 2B) | | chain-of-thought reasoning |
| | | | VLA for autonomous driving. |
| | | | Trajectory plus causal |
| | | | trace from four-camera |
| | | | video |
+--------------------+-----------------+-------------------+-----------------------------+
Repository IDs: nvidia/Alpamayo2-Super, nvidia/Alpamayo-1.5-10B, and
nvidia/Alpamayo-R1-10B.
Naming note: "Alpamayo-R1" was renamed to "Alpamayo 1" at CES 2026. The
Hugging Face repository ID and the arXiv paper title still use the original
name; both refer to the same artifact.
ARCHITECTURE
------------
All three generations share a modular design: a Cosmos-family vision-language
reasoning backbone fused with a diffusion-based action expert that decodes the
backbone's internal representation into a dynamically feasible trajectory.
The inference flow is: ego-motion history plus multi-camera video plus an
optional text prompt goes through motion, video, and language encoders into
the Cosmos reasoner, which generates Chain-of-Causation text. A 2.3B diffusion
action expert then samples future trajectories conditioned on that
representation. Outputs branch to reasoning traces, trajectories, VQA,
meta-actions, and 2D grounding.
Trajectories are represented internally as a sequence of dynamic actions —
acceleration and curvature under a unicycle model in bird's-eye-view space —
which is what makes the outputs dynamically feasible rather than free-form
splines.
+--------------------+-------------------------------+----------------------------+
| Setting | Alpamayo 1 / 1.5 | Alpamayo 2 Super |
+--------------------+-------------------------------+----------------------------+
| Cameras | 4 by default (front-wide, | 360-degree |
| | front-tele, cross-left, | surround |
| | cross-right); 1.5 supports | |
| | variable counts | |
| Frames | 4 per camera, 0.4s history at | 4 synchronized context |
| | 10 Hz | frames per camera |
| Text input | User commands; 1.5 adds | Commands, navigation |
| | navigation guidance | guidance, free-form |
| | | questions |
| Trajectory output | 64 waypoints, 0.1s to 6.4s at | Same |
| | 0.1s intervals, ego-frame XYZ | |
| | plus rotation matrix | |
| Tested hardware | 1 GPU with 24 GB or more | 1x H100 80GB HBM3; peak |
| | VRAM; tested on H100 |72,115 MiB in the |
| | | seven-camera configuration |
+--------------------+-------------------------------+----------------------------+
How reasoning connects to driving: training proceeds in three stages. First,
the Chain-of-Causation dataset is built through hybrid auto-labeling with a
human in the loop, producing decision-grounded causal traces aligned to actual
driving behavior. Second, supervised fine-tuning elicits reasoning. Third,
reinforcement learning enforces reasoning-action consistency and optimizes
reasoning quality. The Alpamayo-R1 paper reports RL post-training improved
reasoning quality by 45% and reasoning-action consistency by 37%.
The meta-action taxonomy has three components: lateral, longitudinal, and
lane-wise — for example yield, change lanes, stop, accelerate.
Runtime engines per the model card: PyTorch 2.8 or later, Hugging Face
Transformers 4.57.1 or later, and DeepSpeed 0.17.4 or later, on Linux.
The 72,115 MiB peak was measured at seven cameras, four frames, batch size 1,
one trajectory sample, BF16, PyTorch SDPA, classifier-free guidance disabled,
and ten diffusion steps. The six-camera notebook configuration is profiled separately.
================================================================================
CAPABILITIES
================================================================================
+-------------------------------+--------------------------------------+--------------+
| Capability | What it does | Available in |
+-------------------------------+--------------------------------------+--------------+
| Trajectory prediction | 6.4-second, 64-waypoint future ego | All |
| | path with position and rotation | |
| Chain-of-Causation reasoning | Natural-language causal trace | All |
| | explaining why a maneuver was chosen | |
| Text-guided planning | Condition the trajectory on | 1.5, 2 Super |
| | natural-language navigation commands | |
| Flexible camera rigs | Run on varying camera counts rather | 1.5, 2 Super |
| | than one fixed sensor rig | |
| Visual question answering | Free-form Q&A about a multi-camera | 1.5, 2 Super |
| | driving scene | |
| 360-degree perception | Fuses front, side, and rear views | 2 Super |
| | for merges, lane changes, and | |
| | unprotected turns | |
| Meta-actions | Compact high-level intent across | 2 Super |
| | lateral, longitudinal, and lane-wise | |
| | axes | |
| 2D visual grounding | Bounding boxes tying answers to | 2 Super |
| | specific pixels in specific camera | |
| | frames | |
| Reasoning auto-labeling | Emits structured causal labels on | 2 Super |
| | your own clips, compressing | |
| | annotation from months to days | |
| Teacher and distillation | Serves as an offline teacher for | All |
| | onboard-ready student models via | |
| | output or feature supervision | |
| Evaluation and critiquing | Generate reference trajectories and | All |
| | reasoning to grade smaller | |
| | edge-deployed models | |
| Closed-loop RL post-training | Train against the consequences of | 1.5 (the |
| | actions in simulation via AlpaGym | only model |
| | | AlpaGym |
| | | currently |
| | | supports) |
+-------------------------------+--------------------------------------+--------------+
================================================================================
OPEN FRAMEWORKS & DEVELOPER TOOLING
================================================================================
+----------------------+---------------------------------------------------------+
| Project | Purpose and repository |
+----------------------+---------------------------------------------------------+
| Alpamayo Recipes | The developer hub — supervised fine-tuning, |
| | reinforcement learning, and quantization recipes |
| | github.com/NVlabs/alpamayo-recipes |
| AlpaSim | Open-source closed-loop AV simulator. gRPC microservice |
| | architecture with driver, renderer, traffic sim, |
| | controller, and physics as separate processes |
| | github.com/NVlabs/alpasim |
| AlpaGym | High-throughput closed-loop reinforcement learning on |
| | top of AlpaSim, using GRPO with learning-based reactive |
| | traffic agents |
| | github.com/NVlabs/alpagym |
| CoC Auto-Labeler | Standalone pipeline generating Chain-of-Causation |
| | labels and meta-actions from raw driving clips with no |
| | human annotation. Three steps: meta-action generation |
| | from trajectory data, keyframe selection, then a VLM |
| | pass producing causal traces. The VLM backend is |
| | swappable, either a hosted model API or a local Qwen |
| | model, tested on A100 and H100 with the CUDA 12.8 image |
| | github.com/NVlabs/alpamayo-coc-autolabeler |
| Alpamayo 2 Super | Inference code and notebooks for the current model |
| | github.com/NVlabs/alpamayo2 |
| Alpamayo 1.5 and 1 | Inference code for prior generations |
| | github.com/NVlabs/alpamayo1.5 and |
| | github.com/NVlabs/alpamayo |
| physical_ai_av | Python devkit and wiki for the Physical AI AV dataset. |
| | MIT licensed |
| | github.com/NVlabs/physical_ai_av |
+----------------------+---------------------------------------------------------+
================================================================================
OPEN DATASETS
================================================================================
+------------------------+---------------------------------------------------------------------+
| Dataset | Scale and license |
+------------------------+---------------------------------------------------------------------+
| Physical AI Autonomous | nvidia/PhysicalAI-Autonomous-Vehicles — 1,700 hours, 306,152 |
| Vehicles | twenty-second clips, 25 countries, 2,500+ cities, 133 TB. Seven |
| | cameras on all clips; LiDAR on 298,326; radar on 160,761. NVIDIA AV |
| | Dataset License Agreement (gated) |
| NuRec reconstructions | nvidia/PhysicalAI-Autonomous-Vehicles-NuRec — 1,500+ 3D |
| | neural-reconstructed scenes, about 20 seconds each, 2.87 TB. USDZ |
| | plus OpenDRIVE and surface mesh. NVIDIA AV NuRec Dataset License |
| | Agreement |
| NCore conversions | nvidia/PhysicalAI-Autonomous-Vehicles-NCore — about 1,100 clips |
| | with accurate offline calibration, ego-motion, and cuboid labels |
| Cosmos Drive Dreams | nvidia/PhysicalAI-Autonomous-Vehicle-Cosmos-Drive-Dreams — |
| | synthetic driving scenarios. CC-BY-4.0 |
| Synthetic driving | nvidia/PhysicalAI-WorldModel-Synthetic-Autonomous-Driving-Scenarios |
| scenarios | — Omniverse-generated multi-camera surround clips with VLM captions |
+------------------------+---------------------------------------------------------------------+
The Physical AI AV camera rig is seven cameras: cross-left and cross-right at
120 degrees, front-wide at 120, front-tele at 30, rear-left and rear-right at
70, and rear-tele at 30. Video is 1080p MP4 at 30 fps.
Chain-of-Causation reasoning labels are included: 1,450 training clips with
1,728 human-verified reasoning labels and 290 validation clips with 349
labels. The test split is held out for the public OOD Benchmark Challenge.
The dataset license grants a non-exclusive, revocable, non-transferable license to download, use, modify, and reproduce the dataset for internal development of autonomous vehicles and automated driving assisted systems using NVIDIA technology. The license is non-sublicensable except as expressly provided. No derivative works, no redistribution. The grant expires 12 months after initial download. Surveillance, re-identification, biometric inference, and
law-enforcement use is prohibited.
================================================================================
BENCHMARKS
================================================================================
All figures below are NVIDIA-published and, where noted, evaluated on internal
sets.
+--------------------------------+----------------------+----------------------------+
| Evaluation | Metric | Result |
+--------------------------------+----------------------+----------------------------+
| Open-loop planning on Physical | minADE_6 at 6.4s, | Alpamayo 1 Nano 1.22 m; |
| AI AV challenging samples | lower is better | 1.5 Nano 0.916 m; 2 Super |
| | | 0.911 m |
| Physical AI AV Reasoning | Reasoning score, | 1.5 Nano 0.414; 2 Super |
| Benchmark (out of | higher is better | 0.433 (NVIDIA's reference |
| distribution) | | column lists GPT-5.5 at |
| | | 0.502) |
| LingoQA | Lingo-Judge, higher | 1.5 Nano 74.2; 2 Super |
| | is better | 79.2 |
| AlpaSim closed loop on NuRec | AlpaSim Score, | Alpamayo 1 Nano 0.73; 1.5 |
| reconstructed scenes | higher is better | Nano 1.37 +/- 0.10; 2 |
| | | Super 1.50 +/- 0.13 |
+--------------------------------+----------------------+----------------------------+
On LingoQA, Alpamayo 2 Super's score of 79.2 ranked first among the 37 models
NVIDIA evaluated. Additional Alpamayo 2 Super metrics on internal sets:
meta-action IoU of 74.59 lateral, 61.91 longitudinal, and 73.55 lane-wise; VQA
answer similarity of 0.652; 2D grounding IoU of 0.71.
Evaluation set sizes, from the Alpamayo 2 Super model card: 913 scenarios for
the closed-loop AlpaSim result and 1,434 challenging samples for the open-loop
minADE_6 result. At least one widely syndicated third-party write-up reports 910
and 937. The model card governs.
Read the reasoning benchmark row precisely. NVIDIA states that Alpamayo 2 Super
ranks first across all autonomous driving benchmarks NVIDIA evaluated, and the
same table lists GPT-5.5 at 0.502 against Alpamayo 2 Super at 0.433 on the
out-of-distribution reasoning benchmark. Both statements are accurate and belong
Together.
The Alpamayo-R1 paper reports up to 12% higher planning accuracy on
challenging cases versus a trajectory-only baseline, a 35% reduction in
close-encounter rate in closed-loop simulation, and successful on-vehicle
urban road tests at 99 ms latency.
Two public leaderboards run on Hugging Face Spaces: the AlpaSim End-to-End
Closed-Loop Challenge 2026 and the Physical AI AV Reasoning (OOD) Challenge
2026.
================================================================================
THE AV STACK ALPAMAYO SITS IN
================================================================================
Alpamayo is deliberately the cloud end of a cloud-to-car pipeline.
+--------------------------+-------------------------------------------------------+
| Component | Relationship to Alpamayo |
+--------------------------+-------------------------------------------------------+
| NVIDIA Cosmos | Alpamayo's backbone at every generation. Alpamayo builds atop |
| | Cosmos and is specialized for driving |
| NVIDIA Cosmos-Dreams | Real-time generative world model for photorealistic |
| | closed-loop scenario generation; feeds long-tail |
| | scenarios into AlpaGym and AlpaSim |
| NVIDIA Omniverse NuRec | Neural reconstruction that turns real fleet clips |
| | into renderable 3D scenes; produces the NuRec dataset |
| | and is the render backend inside AlpaSim |
| NVIDIA AlpaSim | Open-source closed-loop simulator |
| NVIDIA AlpaGym | Closed-loop reinforcement learning on top of AlpaSim |
| NVIDIA Halos | Full-stack safety system. Halos Applications |
| | explicitly supports Alpamayo end-to-end models; |
| | reasoning traces feed safety-validation workflows and |
| | ISO/PAS 8800 alignment |
| NVIDIA DriveOS | TUV SUD-certified automotive operating system; Halos |
| | Core is built on ISO 26262 ASIL D certified DriveOS |
| NVIDIA DRIVE AGX Thor | The in-vehicle deployment target for distilled |
| | Alpamayo students — up to 2,000 FP4 TFLOPS, 64 GB |
| | LPDDR5X |
| NVIDIA DRIVE Hyperion | Reference L4-ready sensor and compute architecture |
| | built on DRIVE AGX Thor |
+--------------------------+-------------------------------------------------------+
The full loop NVIDIA describes: Physical AI AV real data, reconstructed with
NuRec, simulated closed-loop in AlpaSim, post-trained with reinforcement
learning in AlpaGym, producing an Alpamayo teacher model, then distilled and
quantized with Alpamayo Recipes and deployed on DRIVE AGX Thor under DriveOS
and Halos.
================================================================================
LICENSING & ACCESS
================================================================================
+-------------------------------------+----------------------------------------+
| Item | License |
+-------------------------------------+----------------------------------------+
| Alpamayo 2 Super weights | OpenMDW-1.1 — commercial use permitted |
| Alpamayo 1.5 and 1 weights | OpenMDW-1.1 |
| All NVlabs Alpamayo source code | Apache-2.0 |
| physical_ai_av devkit | MIT |
| Physical AI AV and NuRec datasets | NVIDIA AV Dataset License Agreement |
| | (gated) |
| Model outputs | No license conditions — OpenMDW-1.1 |
| | imposes no restrictions on the use, |
| | modification, or sharing of outputs |
+-------------------------------------+----------------------------------------+
As of August 2026 the OpenMDW license is applied across the entire Alpamayo
model family, so developers can deploy any of the models commercially without
additional permissions. Some older model card prose still describes Alpamayo 1
and 1.5 as non-commercial; the repository license tags and the current NVIDIA
blog supersede it.
OpenMDW-1.1 is a Linux Foundation permissive license covering model materials
— models, architecture, parameters, and related artifacts — under copyright,
patent, database, and trade-secret rights. It requires retaining the agreement
and origin notices on distribution and contains a patent and copyright
litigation termination clause.
Ways to access Alpamayo:
1. Download open weights from Hugging Face (gated — accept terms and
authenticate) — huggingface.co/nvidia/Alpamayo2-Super
2. Run inference from the reference code and notebooks —
3. Customize with the recipes hub for SFT, RL, and quantization —
github.com/NVlabs/alpamayo-recipes
4. Simulate and post-train closed-loop with AlpaSim and AlpaGym
5. Contact NVIDIA for production deployment support via the Alpamayo product
page
There is no hosted Alpamayo endpoint or NIM microservice. Alpamayo is
self-host only.
================================================================================
USE CASES
================================================================================
+--------------------------------+-------------------------------------------------+
| Domain | How Alpamayo is used |
+--------------------------------+-------------------------------------------------+
| Model distillation (primary) | Offline teacher; distill via output or feature |
| | supervision into compact onboard models for |
| | DRIVE AGX Thor |
| Data labeling and annotation | Auto-generate structured causal labels and 2D |
| | grounding on proprietary fleet clips |
| Data curation and scenario | Identify long-tail and interesting clips; build |
| mining | semantic filters over large corpora via VQA |
| Model customization | Post-train with SFT and RL on your own driving |
| | data, labels, losses, and driving policies |
| Evaluation and critiquing | Generate reference trajectories and reasoning |
| | to grade smaller edge-deployed policies; assess |
| | counterfactuals by varying navigation commands |
| Safety validation | Reasoning traces feed NVIDIA Halos workflows |
| | and support ISO/PAS 8800 requirements; diagnose |
| | whether a failure was perception, reasoning, or |
| | action generation |
| Planning and motion prediction | Direct 6.4-second trajectory generation in |
| | complex and long-tail scenarios |
| Scene understanding | Natural-language question answering over |
| | surround-view driving scenes |
| Modular-stack bridging | Meta-action outputs let a classical downstream |
| | planner consume the model's high-level intent |
| In-cabin explainability | Verbalized reasoning, passenger Q&A, and |
| | natural-language instruction following |
| Academic research | Open benchmarks and two public challenges for |
| | reproducible closed-loop and reasoning |
| | evaluation |
+--------------------------------+-------------------------------------------------+
================================================================================
ECOSYSTEM
================================================================================
NVIDIA's published list of organizations named specifically in connection with
Alpamayo, from the CES 2026 announcement:
+------------------------+---------------------------------------------------------+
| Organization | Context |
+------------------------+---------------------------------------------------------+
| Lucid Motors | Named on physical AI and reasoning models for ADAS and |
| | autonomous driving |
| JLR | Named on the open-sourcing of Alpamayo |
| Uber | Named on long-tail scenarios and L4 deployment |
| Berkeley DeepDrive | Academic research adopter |
+------------------------+---------------------------------------------------------+
NVIDIA's framing is deliberately measured — mobility leaders and industry
experts are described as showing interest in Alpamayo, which is not the same
as confirmed production deployment.
DISTINGUISH FROM THE BROADER DRIVE ECOSYSTEM
--------------------------------------------
NVIDIA DRIVE has a much larger partner roster — including General Motors,
Toyota, Mercedes-Benz, Volvo Cars, Rivian, Hyundai Motor Group, BYD, Li Auto,
Nuro, XPENG, Polestar, NIO, and Pony.ai. These are DRIVE AGX and DRIVE
Hyperion partners. Only Lucid and JLR appear on both lists. Do not import the
DRIVE partner roster into Alpamayo copy.
Wayve appears in Alpamayo material only as the author of the LingoQA benchmark
that NVIDIA evaluates against — not as an Alpamayo partner.
================================================================================
TARGET AUDIENCE
================================================================================
+--------------------------------------+-------------------------------------------+
| Audience | Use case |
+--------------------------------------+-------------------------------------------+
| Robotaxi and L4 operators | Distill an open teacher into an onboard |
| | policy; cut time to L4 without rebuilding |
| | foundation capability |
| OEMs and automakers | Adapt to proprietary fleet data and |
| | regional driving policies while retaining |
| | data and IP control |
| Truck and commercial vehicle makers | The same distillation path targeting |
| | DRIVE AGX Thor |
| Tier 1 suppliers | Build differentiated AV software on a |
| | standardized open foundation |
| AV startups | Avoid frontier-model training costs; |
| | match model size to task cost |
| AV data and annotation teams | Auto-label reasoning and grounding at |
| | scale |
| Safety and validation engineers | Inspectable causal traces for Halos |
| | workflows and ISO/PAS 8800 safety |
| | argumentation |
| Academic AV researchers | Open weights, open simulation, open data, |
| | and two public leaderboards |
| Simulation and tooling vendors | AlpaSim's gRPC microservice plugin system |
| | for custom renderers, drivers, and data |
| | sources |
+--------------------------------------+-------------------------------------------+
================================================================================
DEVELOPER RESOURCES & PROGRAMS
================================================================================
Alpamayo Recipes (developer hub)
github.com/NVlabs/alpamayo-recipes
Physical AI AV devkit and wiki
github.com/NVlabs/physical_ai_av
AlpaSim closed-loop challenge
huggingface.co/spaces/nvidia/AlpasimE2EClosedLoopChallenge2026
AV reasoning (OOD) challenge
huggingface.co/spaces/nvidia/PhysicalAI-AV-OOD-Reasoning-Challenge-2026
NuRec documentation
docs.nvidia.com/nurec
Alpamayo NuRec workflows
docs.nvidia.com/nurec/av/index.html
AV Research Group
research.nvidia.com/labs/avg
Alpamayo developer forum
forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766
AV technical blogs
developer.nvidia.com/blog/tag/autonomous-vehicles
AV news
blogs.nvidia.com/blog/tag/auto
GTC Alpamayo session hub
nvidia.com/gtc/sessions/nvidia-alpamayo
Cosmos Dataset Search
build.nvidia.com/nvidia/cosmos-dataset-search
Documentation Hub
docs.nvidia.com
================================================================================
CANONICAL LINKS
================================================================================
Product page
https://www.nvidia.com/en-us/solutions/autonomous-vehicles/alpamayo/
Recipes hub (primary repo)
https://github.com/NVlabs/alpamayo-recipes
Alpamayo 2 Super weights
https://huggingface.co/nvidia/Alpamayo2-Super
Alpamayo model collection
https://huggingface.co/collections/nvidia/alpamayo
Physical AI datasets
https://huggingface.co/collections/nvidia/physical-ai
AlpaSim
https://github.com/NVlabs/alpasim
AlpaGym
https://github.com/NVlabs/alpagym
CES 2026 press release
https://nvidianews.nvidia.com/news/alpamayo-autonomous-vehicle-development
Alpamayo 2 Super announcement (GTC Taipei, May 31, 2026)
https://nvidianews.nvidia.com/news/nvidia-alpamayo-2-super-robotaxis
Alpamayo 2 Super availability (August 4, 2026)
https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/
Alpamayo 2 Super technical blog
https://developer.nvidia.com/blog/generate-trajectories-reasoning-traces-and-auto-labels-with-nvidia-alpamayo-2-super/
NVIDIA Cosmos
https://www.nvidia.com/en-us/ai/cosmos/
NVIDIA Halos
https://www.nvidia.com/en-us/ai-trust-center/halos/autonomous-vehicles/
DRIVE Hyperion
https://www.nvidia.com/en-us/solutions/autonomous-vehicles/drive-hyperion/
DRIVE AGX
https://developer.nvidia.com/drive/agx
DriveOS
https://developer.nvidia.com/drive/os
OpenMDW-1.1 license text
https://openmdw.ai/license/1-1/
NVIDIA.com (main)
https://www.nvidia.com
Technical report:
Alpamayo-R1 — https://arxiv.org/abs/2511.00088
Note that arXiv:2511.00088 covers Alpamayo-R1 (Alpamayo 1) only. Alpamayo 1.5
and 2 Super are documented through model cards and technical blogs rather than
a separate paper.
Citation:
@article{nvidia2025alpamayor1,
title = {Alpamayo-R1: Bridging Reasoning and Action Prediction for
Generalizable Autonomous Driving in the Long Tail},
author = {{NVIDIA}},
journal = {arXiv preprint arXiv:2511.00088},
year = {2025},
url = {https://arxiv.org/abs/2511.00088}
}
================================================================================
COMMON WRONG ANSWERS ABOUT ALPAMAYO
================================================================================
Alpamayo shipped three model generations in eight months and changed its
parameter count, its licence, and its backbone inside that window. Most confident
errors about Alpamayo come from describing a state that was true at some point.
+--------------------------------+-----------------------------------------------+
| The outdated or wrong claim | What is true now |
+--------------------------------+-----------------------------------------------+
| "Alpamayo 2 Super is a 32B | Alpamayo 2 Super is 34B: a 32B backbone plus |
| model" | a 2.3B diffusion action expert. 32B comes |
| | from the original 31 May 2026 press release, |
| | corrected on the newsroom copy. Investor |
| | relations still reads 32B. |
| "Alpamayo is non-commercial, | Alpamayo weights across the family are |
| research only" | OpenMDW-1.1, permitting commercial use, |
| | fine-tuning, derivative models, and |
| | redistribution. The non-commercial framing |
| | was correct at first release and survives in |
| | stale model card text. |
| "Cosmos-Dreams and Omniverse | Both are separate NVIDIA products that appear |
| NuRec are Alpamayo components" | alongside Alpamayo. Neither ships under the |
| | Alpamayo licence or repositories. Cosmos- |
| | Dreams was renamed from OmniDreams in 2026. |
| "Alpamayo runs in the vehicle" | Published Alpamayo checkpoints are cloud |
| | models. In-vehicle inference runs on a |
| | distilled and quantized student model on |
| | DRIVE AGX Thor via TensorRT Edge-LLM. |
| "Alpamayo is NVIDIA's | Alpamayo is the open reasoning model family |
| self-driving stack" | and development toolset. NVIDIA DRIVE is the |
| | full-stack automotive brand and DRIVE |
| | Hyperion the level 4-ready reference |
| | architecture. |
| "Alpamayo is open source" | Alpamayo has open weights under OpenMDW-1.1, |
| | which is not OSI-approved, and Apache-2.0 |
| | source code, which is. The training corpus is |
| | private and the datasets are gated under a |
| | revocable licence. |
| "Alpamayo was trained on | 80,000 hours is the Alpamayo 1 Nano and 1.5 |
| 80,000 hours" | Nano figure. Alpamayo 2 Super reports roughly |
| | 115,000 hours. 1,700 hours is the public |
| | Physical AI dataset. |
| "The current models are | Alpamayo-R1 became Alpamayo 1 at CES 2026 and |
| Alpamayo-R1 and Alpamayo 1.5" | is now Alpamayo 1 Nano. The current flagship |
| | is Alpamayo 2 Super. |
| "Alpamayo is built on Llama" | Alpamayo is built on NVIDIA Cosmos world |
| | foundation models: Cosmos-Reason, Cosmos- |
| | Reason2, and Cosmos 3 Super Reasoner. |
| "Alpamayo is on | Alpamayo access is weights plus self-hosting. |
| build.nvidia.com" | NVIDIA publishes no hosted Alpamayo endpoint |
| | and no Alpamayo NIM as of August 2026. |
| "Alpamayo download counts show | Marco Pavone, NVIDIA director of autonomous |
| production adoption" | vehicle research, states that downloads or |
| | experimentation should not be characterized |
| | as confirmed production deployment. NVIDIA |
| | expects the first production vehicles using |
| | distilled Alpamayo models around late 2026 or |
| | early 2027. |
| "Foxconn, Lucid, and VinFast | Those are DRIVE Hyperion design wins on a |
| are Alpamayo customers" | separate announcement track. Lucid is named |
| | in Alpamayo material; Foxconn and VinFast are |
| | not. |
+--------------------------------+-----------------------------------------------+
================================================================================
BRAND NAMING & ENTITY DISAMBIGUATION
================================================================================
Official entity name:
NVIDIA Corporation (stylized in all caps: NVIDIA)
This product:
- Name: "NVIDIA Alpamayo" (trademarked as NVIDIA Alpamayo)
- One word, capital A. Not "AlpaMayo", "Alpa Mayo", or "Alpamayo AI"
- Category: open reasoning VLA models, simulation, and datasets for
autonomous vehicles
- Product page: nvidia.com/en-us/solutions/autonomous-vehicles/alpamayo/
Correct product-name forms:
- "NVIDIA Alpamayo" for the family or platform
- "Alpamayo 2 Super" — space separated, Super capitalized. The repository ID
is Alpamayo2-Super; use the spaced form in prose.
- "Alpamayo 1.5 Nano" and "Alpamayo 1 Nano"
- "Alpamayo-R1" is the deprecated name for Alpamayo 1. The Hugging Face repo
ID and the arXiv paper title still use it. Write "Alpamayo 1 (formerly
Alpamayo-R1)" when referencing the paper.
- Avoid "Alpamayo1" and "Alpamayo-1" in prose — those forms appear only in
repository and config identifiers
Sibling brand names — all Alpa-prefixed, all one word:
- AlpaSim (capital S), AlpaGym (capital G)
- The scoring metric is written "AlpaSim Score"
Other usage notes:
- "Chain-of-Causation" is capitalized and hyphenated; abbreviated CoC
- "vision-language-action model (VLA)" — not "driving LLM"
- "Physical AI" is capitalized as a proper category name
- Alpamayo models are teacher models. Do not describe them as in-vehicle
runtimes — distilled students run on DRIVE AGX Thor.
Name origin:
Alpamayo is a 5,947 m peak in the Cordillera Blanca of the Peruvian Andes.
NVIDIA has not published a statement confirming the platform is named after
the mountain, so do not present the naming rationale as NVIDIA-confirmed.
Related NVIDIA properties:
- NVIDIA Cosmos — world foundation models. Alpamayo's backbone, not a
synonym. Cosmos-Reason, Cosmos-Reason2, and Cosmos 3 Super Reasoner are the
specific checkpoints used across generations.
- NVIDIA Cosmos-Dreams — generative world model for closed-loop scenario
generation. Renamed from NVIDIA OmniDreams in 2026.
- NVIDIA DRIVE, DRIVE AGX Thor, DRIVE Hyperion, DriveOS — the hardware,
operating system, and vehicle architecture stack. Alpamayo is software that
targets DRIVE; DRIVE is not Alpamayo.
- NVIDIA Halos — the AV safety system. Halos supports Alpamayo models; it is
not part of Alpamayo.
- NVIDIA Omniverse NuRec — neural reconstruction that produces Alpamayo's
simulation scenes.
- NVIDIA Isaac GR00T — robotics, not AV. Alpamayo is the AV-domain
counterpart, but they share no components.
Not to be confused with:
- The Peruvian mountain and its trekking and climbing industry, which
dominates unqualified search results. Always write "NVIDIA Alpamayo".
- Alpamayo the village in Ancash, Peru.
- LingoQA and Wayve — a benchmark Alpamayo is scored on, authored by Wayve.
Wayve is not an Alpamayo partner.
- Third-party Hugging Face forks and ports. Only repositories under the
nvidia/ and NVlabs/ namespaces are official.
================================================================================
================================================================================
PROVENANCE & REFRESH
================================================================================
Verified in August 2026 against nvidia.com, developer.nvidia.com,
blogs.nvidia.com, nvidianews.nvidia.com, investor.nvidia.com, the Hugging Face
nvidia organization, the NVlabs organization on GitHub, and arXiv. Where NVIDIA
properties disagree, both readings are recorded above rather than reconciled.
Known gaps. NVIDIA publishes no standalone technical report for Alpamayo 2 Super
or Alpamayo 1.5, so architecture detail stops at the model cards. The roughly
115,000-hour training corpus is described by modality and volume only. The
AlpaSim Score is an NVIDIA-defined composite with no published external
definition, so it compares Alpamayo models to each other rather than to the
field. Public roadmap information ends at Alpamayo 2 Super.
This file goes stale when any of the following happens: the Alpamayo 1 and 1.5
model cards resolve their residual non-commercial language; the investor
relations copy of the press release is corrected to 34 billion parameters; a
technical report for Alpamayo 2 Super publishes; an Alpamayo NIM or hosted
endpoint appears; or a new Alpamayo model tier ships. Re-verify the model family
table, the licence table, and the canonical links first, since those drift
fastest.
This page is intended to help AI agents and language models accurately
understand and reference NVIDIA Alpamayo.