AI Wiki

NVIDIA Alpamayo 2 Super

AI ModelsAutonomous VehiclesNVIDIA
15 min read
Updated
Suggest editHistoryTalk17 citationsFact-checked
RawGraph

NVIDIA Alpamayo 2 Super is an open, 34-billion-parameter reasoning-based vision-language-action model (VLA) for safe, Level 4 robotaxi and autonomous-vehicle development from NVIDIA. Announced at GTC Taipei during COMPUTEX 2026 on June 1, 2026, the model was released for commercial use on August 4, 2026, with open weights published on Hugging Face under the OpenMDW-1.1 license and inference code on GitHub under the Apache License 2.0.[1][2][7] It is the flagship member of the open Alpamayo family, which NVIDIA first introduced at CES in January 2026, and combines a 32-billion-parameter NVIDIA Cosmos 3 Super Reasoner vision-language backbone with a diffusion-based Action Expert (2 billion parameters per NVIDIA developer blog and GitHub documentation, 2.3 billion per the model card parameter listing), post-trained with reinforcement learning.[6][7] The model perceives a vehicle's full 360-degree surroundings across up to seven cameras, reasons step by step about complex driving situations, and explains the decisions it makes; at release, NVIDIA reported that it ranked first on the LingoQA autonomous-driving reasoning benchmark among nearly 40 models evaluated.[2][6] NVIDIA positions it as a "teacher" model that developers can fine-tune and distill into smaller networks that run in real time on in-vehicle hardware.[1][2]

Alpamayo, which shares its name with a peak in the Peruvian Andes, is built on the NVIDIA Cosmos family of world foundation models. Alongside Alpamayo 2 Super, NVIDIA announced a set of companion tools meant to complete the pipeline from real-world data capture to closed-loop training and on-road deployment, including the AlpaGym reinforcement learning framework, the Cosmos-Dreams generative world model (announced as OmniDreams and later renamed), and new NVIDIA Omniverse NuRec neural-reconstruction models.[1][3]

Key facts

AttributeDetail
Full nameNVIDIA Alpamayo 2 Super
DeveloperNVIDIA
AnnouncedJune 1, 2026, at GTC Taipei (COMPUTEX 2026)
ReleasedAugust 4, 2026, for commercial use
Model typeReasoning-based vision-language-action (VLA) model
ParametersAbout 34 billion nominal: 32-billion-parameter VLM backbone plus a diffusion Action Expert given as 2 billion (developer blog, GitHub) or 2.3 billion (model card) parameters; announced in June 2026 as 32 billion
FoundationNVIDIA Cosmos 3 Super Reasoner, post-trained with reinforcement learning
Weightshuggingface.co/nvidia/Alpamayo2-Super, OpenMDW-1.1 license
Codegithub.com/NVlabs/alpamayo2 (inference), Apache License 2.0
BenchmarksFirst on LingoQA among nearly 40 models (Lingo-Judge score 79.2)
Target useLevel 4 robotaxi and autonomous-vehicle development
PredecessorsAlpamayo 1 (CES, January 2026) and Alpamayo 1.5 (March 2026), both 10-billion-parameter class
Companion toolsAlpaGym, Cosmos-Dreams (formerly OmniDreams), Omniverse NuRec, AlpaSim
DownloadsFamily surpassed 500,000 Hugging Face downloads as of August 4, 2026
Deployment targetDistilled onto NVIDIA DRIVE AGX Thor within DRIVE Hyperion

Background and the Alpamayo family

NVIDIA launched the Alpamayo family at CES on January 5, 2026, describing it as an open, large-scale collection of reasoning VLA models, simulation frameworks, and physical-AI datasets for autonomous driving, and called itself the first to release an open reasoning VLA model. The goal was to give the global robotaxi and AV industry a shared, transparent foundation so that developers would not have to build core autonomy infrastructure from scratch.[3][4]

The original Alpamayo 1 is a 10-billion-parameter chain-of-thought, reasoning-based VLA model. Rather than mapping camera pixels directly to steering and braking commands the way a perception-only stack does, Alpamayo takes in multi-camera video, navigation inputs, and driving context, then generates a driving trajectory together with a reasoning trace that explains the choice. NVIDIA calls these traces Chain-of-Causation (CoC) reasoning. The approach lets a vehicle think through a novel or rare scenario step by step; NVIDIA's developer blog cites examples such as construction zones, partially occluded pedestrians, unusual right-of-way interactions, and objects entering the roadway.[3][4][6]

The technical foundation for the family was described in the research paper "Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail," posted to arXiv by an NVIDIA team in late 2025. That work pairs Cosmos-Reason, a vision-language model pre-trained for physical AI, with a diffusion-based trajectory decoder that produces dynamically feasible trajectories in real time. It also introduces the Chain-of-Causation dataset, built through a hybrid auto-labeling and human-in-the-loop pipeline that produces decision-grounded, causally linked reasoning traces aligned with driving behaviors. The paper reported that scaling the model from 0.5 billion to 7 billion parameters produced consistent gains, foreshadowing the much larger Alpamayo 2 Super.[5]

NVIDIA extended the family with Alpamayo 1.5, published on Hugging Face on March 19, 2026. Like Alpamayo 2 Super, it pairs a Cosmos-Reason vision-language backbone (Cosmos-Reason2) with a diffusion-based trajectory decoder and reinforcement learning post-training; its model card describes an 8.2-billion-parameter backbone combined with a 2.3-billion-parameter action expert, keeping it in the same 10-billion-parameter class as Alpamayo 1.[9] NVIDIA describes both earlier models as having been initially introduced for research and development, with Alpamayo 2 Super delivering the family's highest reasoning and driving performance and the smaller models serving as more cost-efficient options for cloud-based development and distillation.[2]

By the time Alpamayo 2 Super was announced, NVIDIA said the Alpamayo models had been downloaded close to 400,000 times, and the lineup won a COMPUTEX 2026 Best Choice Award in the Vehicle Technology and Smart Cockpit category.[1][3] By the August 4, 2026 release, the family had surpassed 500,000 Hugging Face downloads, which NVIDIA describes as making it the most-adopted open reasoning model family for autonomous driving on the platform.[2]

What "Super" denotes

The "Super" suffix marks Alpamayo 2 as the high-capacity, top-of-range member of the second generation, in contrast to the smaller Nano variants NVIDIA offers for on-device use. The most concrete difference is scale: NVIDIA describes Alpamayo 2 Super as offering 3x the scale of the 10-billion-parameter Alpamayo 1 and Alpamayo 1.5 models, with the added capacity helping the model generalize reasoning from sparse examples in rare, multi-agent interactions.[2] NVIDIA frames the larger model as a teacher that captures the broadest possible driving knowledge and reasoning ability, which developers then distill into compact student models small enough to run inside a vehicle.[1][2]

Beyond raw size, Alpamayo 2 Super adds several capabilities aimed squarely at the hardest cases in autonomous driving:

CapabilityWhat it adds
Full-surround perceptionExpands from front-focused cameras to 360-degree situational awareness across front, side, and rear views
Meta-ActionsAdds macro driving actions such as yield, lane change, and stop to the model's output
Reasoning auto-labeling with 2D groundingGenerates high-quality reasoning labels tied to image regions, compressing annotation cycles from months to days
Improved Chain-of-Causation tracesProduces better reasoning and trajectories in rare, complex, long-tail scenarios
3D spatial understandingStrengthens spatial reasoning and trajectory prediction in difficult conditions

The full-surround upgrade is significant because earlier reasoning was concentrated on what the vehicle could see ahead. Extending it to a 360-degree view lets the model reason about merging traffic, vehicles approaching from the side, and events behind the car. The Meta-Actions output gives the model a vocabulary of high-level maneuvers, which makes its intentions easier to interpret and validate. NVIDIA emphasizes that this interpretability is meant to support safety validation and collaboration with regulators, since the model can verbalize why it chose a given action.[1][3]

August 2026 release and licensing

On August 4, 2026, NVIDIA released Alpamayo 2 Super for commercial use. The model weights are published on Hugging Face at nvidia/Alpamayo2-Super under OpenMDW-1.1, which NVIDIA describes as the Linux Foundation's permissive license for open AI model distributions; the license covers fine-tuning, derivative models, and commercial redistribution, allowing AV developers, automakers, truckmakers, and suppliers to adapt the model to their own data, driving policies, and deployment strategies.[2][7] Inference code, including command-line tools, notebooks, and APIs for running the model on NVIDIA Physical AI dataset samples, is available in the NVlabs/alpamayo2 repository on GitHub under the Apache License 2.0.[7][8] NVIDIA said the earlier Alpamayo releases had been initially introduced for research and development, and that the OpenMDW license was being applied across the entire Alpamayo model family so that developers can deploy any of the models commercially without requiring additional permissions.[2]

The released model is built on the NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning. It produces five coupled outputs from the same driving context: a trajectory describing the planned path, a Chain-of-Causation reasoning trace explaining the decision, high-level meta-actions capturing the model's intent, reasoning auto-labels for generating CoC annotations at scale, and visual question answering with 2D grounding.[2][6] The trajectory output covers 64 waypoints spanning up to 6.4 seconds.[7] NVIDIA notes that the model can also be deployed as an auto-labeler on proprietary fleet data, and that its CoC traces integrate with NVIDIA Halos safety-validation workflows and support AI safety aligned with ISO/PAS 8800 requirements.[2]

The model's parameter count was revised between announcement and release. NVIDIA's June 1, 2026 press release, as originally published, described Alpamayo 2 Super as a 32-billion-parameter model.[10] The release-day model card, GitHub repository, and NVIDIA developer blog instead specify 34 billion parameters in total, combining the 32-billion-parameter Cosmos 3 Super Reasoner vision-language backbone with a diffusion-based Action Expert; the model card parameter listing gives the Action Expert as 2.3 billion parameters while the card prose, the GitHub README, and the developer blog round the pair to 34 billion with a 2-billion-parameter expert, and the live version of the June press release now also reads 34 billion.[1][6][7][8]

Benchmark results

NVIDIA reported at release that Alpamayo 2 Super ranks first on LingoQA, an autonomous-driving reasoning benchmark, among nearly 40 models evaluated. In NVIDIA's testing using the Lingo-Judge metric, the model outperformed Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points.[2] The model card and NVIDIA's developer blog report a Lingo-Judge score of 79.2, alongside an open-loop minADE_6 of 0.911 meters at a 6.4-second horizon and a closed-loop AlpaSim score of 1.50 +/- 0.13.[6][7] NVIDIA further states that the model ranks first across all autonomous driving benchmarks the company evaluated; this is a claim based on NVIDIA's own testing rather than an independent leaderboard.[2] NVIDIA's own companion developer blog complicates the superlative: its benchmark table lists OpenAI's GPT-5.5 at 0.502 against Alpamayo 2 Super's 0.433 on NVIDIA's Physical AI AV Reasoning Benchmark, where higher scores are better.[6]

Companion tools and the training pipeline

Alpamayo 2 Super was announced alongside a toolchain that NVIDIA describes as completing the path from real-world data capture to closed-loop training and in-vehicle deployment.[1][3] At the August 2026 release, NVIDIA grouped the model with NVIDIA AlpaSim, NVIDIA AlpaGym, the NVIDIA Physical AI Open Datasets, and open training recipes and an auto-labeling pipeline.[2]

AlpaSim

AlpaSim is NVIDIA's open-source, end-to-end closed-loop simulation framework for autonomous vehicles, first released with the Alpamayo family at CES. It validates model decisions against simulated real-world consequences and provides realistic sensor modeling, configurable traffic dynamics, and scalable closed-loop testing.[3][4]

AlpaGym

AlpaGym is an open-source, high-throughput, closed-loop reinforcement learning framework. It runs models through continuous cycles of decision and observation inside the AlpaSim environment, which exposes compounding errors and edge-case failures that static, pre-recorded datasets tend to miss. Training a driving policy in this loop, rather than only on fixed clips of past driving, is meant to make the model more robust when small mistakes would otherwise accumulate.[1][3]

Cosmos-Dreams (formerly OmniDreams)

Cosmos-Dreams is a generative world model for photorealistic, closed-loop AV scenario generation. It lets developers simulate rare and long-tail driving scenarios at scale, generating the kinds of unusual situations that are hard to capture often enough in real-world fleet data. The tool was announced at GTC Taipei as NVIDIA OmniDreams; NVIDIA later renamed it NVIDIA Cosmos-Dreams and updated the June 1 press release to reflect the new name.[1]

Omniverse NuRec

The NVIDIA Omniverse NuRec models add a neural-reconstruction skill that turns real-world fleet recordings into photorealistic 3D scenes. Those reconstructed scenes can then be replayed and modified in simulation, and adapted across different vehicle sensor configurations, so that data collected by one fleet can be reused to train and test other vehicles.[1][3]

Taken together, the tools give what NVIDIA calls a continuous path from open-loop pretraining on recorded data to closed-loop refinement in simulation. The Chain-of-Causation auto-labeling pipeline is also released as open source, which lets developers generate their own causally grounded reasoning labels rather than annotating them by hand.[1][3]

Fit within NVIDIA's autonomous-vehicle stack

Alpamayo 2 Super sits at the software layer of NVIDIA's broader autonomous-vehicle platform. Because a 34-billion-parameter model is far too large to run directly on a vehicle, NVIDIA's intended workflow is to use Alpamayo 2 Super as a teacher and distill it into compact models that run on NVIDIA DRIVE AGX Thor, the in-car compute platform, within the DRIVE Hyperion reference architecture for sensors and compute. The simulation and reconstruction tooling is built on the AlpaSim microservice simulation stack together with Omniverse NuRec.[1][3]

The model's foundation in NVIDIA Cosmos connects it to NVIDIA's wider physical-AI strategy, since Cosmos models underpin both the reasoning core, via Cosmos Reason in the first generation and the Cosmos 3 Super Reasoner in Alpamayo 2 Super, and the generative simulation used to train and test it. In this way Alpamayo spans NVIDIA's data, simulation, and on-vehicle layers as a single open ecosystem rather than a standalone model.[1][3][5][6]

Reception

Coverage of the June 2026 announcement focused on the model's size and on the open, full-stack nature of the Alpamayo ecosystem. Trade and financial press, including GamesBeat and reporting carried through GlobeNewswire and Yahoo Finance, described Alpamayo 2 Super as a 32-billion-parameter step up from the 10-billion-parameter first generation and highlighted the AlpaGym and OmniDreams tools and the planned summer release on GitHub and Hugging Face.[13][14][15] Automotive World framed the model's shift from trajectory generation to reasoning as a departure from how autonomous driving software has been built, arguing that the open strategy undercuts the case for AV developers building foundational reasoning infrastructure from scratch.[16] The Elec placed the launch in the context of NVIDIA's robotaxi ecosystem announcements at GTC Taipei, including Foxconn's planned 2028 robotaxi launch in Taiwan, VinFast in Southeast Asia, Uber in Munich, and Saudi Arabia's HUMAIN.[17]

In NVIDIA's keynote, founder and CEO Jensen Huang summarized the pitch: "Alpamayo is the moment cars begin to safely reason, not just drive. 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."[1]

At the August 4, 2026 release, Huang posted on X that Alpamayo 2 Super is "our frontier open reasoning model for autonomous vehicles" and a backbone for robotaxis, trucks, shuttles, delivery vans, tractors, and mobile robots, saying NVIDIA was releasing it under OpenMDW-1.1 "so teams can inspect it, fine-tune it and deploy it" because "open models advance safety and security", and that "the next wave of AI is robotics", starting with autonomous vehicles.[11] The Next Web described Alpamayo 2 as "a driving model that reasons out loud", producing a chain of reasoning, a trajectory, and a plain-language account of each decision, and noted that the open release complements NVIDIA's hardware business, since distilled versions are meant to run on NVIDIA's in-vehicle computers; it also cautioned that benchmark scores do not equal real-world driving performance and that the model's memory requirements put local use beyond lightly resourced teams.[12]

See also

References

  1. ^NVIDIA Newsroom. "NVIDIA Launches Alpamayo 2 Super Open Reasoning Model for Robotaxis." May 31, 2026 (later updated). nvidianews.nvidia.com/...lpamayo-2-super-robotaxis
  2. ^Jessica Soares, NVIDIA Blog. "NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use." August 4, 2026. blogs.nvidia.com/...super-open-model-now-available
  3. ^NVIDIA. "NVIDIA Alpamayo: Open AI for Robotaxis and Autonomous Vehicles." nvidia.com/...alpamayo
  4. ^NVIDIA Newsroom. "NVIDIA Announces Alpamayo Family of Open-Source AI Models and Tools to Accelerate Safe, Reasoning-Based Autonomous Vehicle Development." January 5, 2026. nvidianews.nvidia.com/...omous-vehicle-development
  5. ^Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail. arXiv:2511.00088. arxiv.org/...2511.00088
  6. ^Boris Ivanovic and Marco Pavone, NVIDIA Technical Blog. "Generate Trajectories, Reasoning Traces, and Auto-Labels with NVIDIA Alpamayo 2 Super." August 4, 2026. developer.nvidia.com/...th-nvidia-alpamayo-2-super
  7. ^NVIDIA. "nvidia/Alpamayo2-Super." Hugging Face model card. August 4, 2026. huggingface.co/...Alpamayo2-Super
  8. ^NVIDIA. "NVlabs/alpamayo2." GitHub repository. github.com/...alpamayo2
  9. ^NVIDIA. "nvidia/Alpamayo-1.5-10B." Hugging Face model card. March 19, 2026. huggingface.co/...Alpamayo-1.5-10B
  10. ^Internet Archive Wayback Machine. Snapshot of "NVIDIA Launches Alpamayo 2 Super Open Reasoning Model for Robotaxis," June 2, 2026. web.archive.org/...nvidia-alpamayo-2-super-robotaxis
  11. ^Jensen Huang. Post on X. August 4, 2026. x.com/...2084656303046332747
  12. ^Ana-Maria Stanciuc, The Next Web. "Nvidia opens up Alpamayo 2, a driving model that reasons out loud." August 4, 2026. thenextweb.com/...o-2-open-reasoning-driving-model
  13. ^NVIDIA Corporation (via GlobeNewswire). "NVIDIA Launches Alpamayo 2 Super Open Reasoning Model for Robotaxis." June 1, 2026. globenewswire.com/...Reasoning-Model-for-Robotaxis
  14. ^GamesBeat. "Nvidia unveils Alpamayo 2 Super AI model for safer robotaxis." May 31, 2026. gamesbeat.com/...uper-ai-model-for-safer-robotaxis
  15. ^NVIDIA Investor Relations. "NVIDIA Launches Alpamayo 2 Super Open Reasoning Model for Robotaxis." June 1, 2026. investor.nvidia.com/...default
  16. ^Automotive World. "Nvidia launches Alpamayo 2 Super open reasoning model." June 1, 2026. automotiveworld.com/...-super-open-reasoning-model
  17. ^The Elec. "Nvidia Unveils 'Alpamayo 2 Super' Open Model for Level 4 Robotaxis." June 2, 2026. thelec.net/...articleView

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

3 revisions · v4 · 2,901 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check of the August 4, 2026 open-release update against NVIDIA's release blog, developer blog, Hugging Face model card, GitHub repository, June and CES press releases (including Wayback snapshots of both), and press coverage. Corrected in this pass: the action-expert parameter figure (NVIDIA's own documents give both 2B and 2.3B; the article now reports the discrepancy), a relayed first-on-all-benchmarks claim that NVIDIA's own developer-blog table contradicts (GPT-5.5 0.502 vs 0.433), the COMPUTEX award category name, the CES family description, and two anecdotes not found in their cited sources. The article's report that NVIDIA silently changed the June press release from 32 to 34 billion parameters was verified against both snapshots.

Cite this page: AI Wiki. "NVIDIA Alpamayo 2 Super." aiwiki.ai, updated 2 Jun 2026, fact-checked 4 Aug 2026. CC BY 4.0. https://aiwiki.ai/wiki/nvidia_alpamayo_2_super

Suggest edit