NVIDIA is releasing Alpamayo 2 Super, an open-source AI reasoning model that is licensed for commercial robotaxi and AV development.
Most autonomous vehicle failures happen in the edge cases: the unprotected left turn with a cyclist cutting through, the four-way merge where nobody has right of way, the delivery truck double-parked around a blind curve.
Those are the situations that resist plain object detection and motion prediction, because handling them requires a vehicle to understand context, weigh cause and effect, and choose an action it can then turn into a path that’s both safe and comfortable.
Doing that in real-time, and in a way that engineers can inspect and validate afterwards, is the problem that NVIDIA’s latest release targets.
NVIDIA has made Alpamayo 2 Super available now for commercial use on Hugging Face. It’s part of the Alpamayo family, which NVIDIA describes as the most-adopted open reasoning models for autonomous driving on the platform.
The new release is built on NVIDIA’s Cosmos 3 Super Reasoner architecture, then post-trained with reinforcement learning. One foundation model now covers a range of AV-relevant tasks instead of requiring separate systems for each one.
Licensing terms that open a path to production
Alpamayo 2 Super ships under OpenMDW-1.1, the Linux Foundation’s permissive licence for open AI model distribution. The terms cover fine-tuning, derivative models, and commercial redistribution. Automakers, truckmakers, and suppliers can adapt the model to their own data and driving policies without negotiating separate commercial terms afterward.
Earlier releases in the Alpamayo family were introduced mainly for research. NVIDIA is now applying OpenMDW-1.1 across the whole family, so developers can move from adaptation to deployment without seeking additional permissions.
Open weights make that transition financially workable, too. Teams can build on advanced reasoning without retraining every foundation capability from scratch, and they don’t have to pay frontier-model prices for tasks that don’t need frontier-model reasoning. Match the model to the job, in other words, and the cost follows.
Within the family, Alpamayo 2 Super does the heaviest reasoning work in cloud-based development: generating reasoning traces, synthetic training data, and teacher outputs used to distil smaller models. Alpamayo 1.5 and Alpamayo 1 sit below it, offering cheaper options for the same cloud workflows.
The resulting distilled models then get optimised for real-time inference inside production vehicles, giving AV programmes frontier-scale reasoning in the cloud and smaller, specialised models on the road.
Benchmark claims from NVIDIA’s testing
NVIDIA’s testing puts Alpamayo 2 Super first on LingoQA, a reasoning benchmark for autonomous driving, among roughly 40 models evaluated.
Using the Lingo-Judge metric, the company recorded a 17.0-point lead over Qwen2.5-VL 72B and a 15.1-point lead over Gemini 2.5 Pro. Against GPT-4o, the margin widened to 23.2 points, by NVIDIA’s own reporting. NVIDIA also puts Alpamayo 2 Super first across every autonomous driving benchmark it evaluated internally.
The model runs at three times the parameter count of the 10-billion-parameter Alpamayo 1.5 and Alpamayo 1 models. NVIDIA says the added capacity helps it generalise reasoning from sparse examples, the rare multi-agent interactions where conventional planning systems tend to struggle.
It also reasons over full-surround camera coverage, fusing front, side, and rear views into a single 360-degree picture. NVIDIA says that fused view improves handling of lane changes, merges, unprotected turns, and complex intersections.
Five outputs from a single model
For every driving situation, Alpamayo 2 Super produces five outputs at once.
There’s a trajectory, the vehicle’s planned path. There’s a chain-of-causation trace explaining the reasoning behind it. A meta-action captures intent – yield, change lanes, stop – in a form that developers can read directly. The model also generates reasoning auto-labels for training data, plus visual question-answering responses tied to specific regions of the camera image through 2D grounding.
Tying those five outputs together lets developers connect what the model observed to the action it chose. That link makes decisions easier to inspect and critique after the fact. The chain-of-causation traces integrate with NVIDIA’s Halos safety-validation workflows and are built to support AI safety practices aligned with ISO/PAS 8800 requirements.
Autolabeling and the wider Alpamayo toolset
Alpamayo 2 Super also works as an autolabeler, applying chain-of-causation labels and 2D-grounded visual question answering to a company’s own fleet footage. NVIDIA says this can turn raw driving clips into training data without months of manual annotation.
Beyond labelling, the model supports scene understanding, model critiquing, and knowledge distillation, so one foundation model can cover more of the development stack instead of requiring separate purpose-built systems for each task.
Alpamayo 2 Super doesn’t stand alone. NVIDIA AlpaSim runs closed-loop simulation. AlpaGym handles high-throughput reinforcement learning. NVIDIA’s Physical AI Open Datasets supply training and testing data, alongside open training recipes and an autolabeling pipeline meant to speed up development and validation cycles.
The Alpamayo family has passed 500,000 downloads on Hugging Face, a figure NVIDIA cites as evidence of its position as the most downloaded open reasoning model family for autonomous driving on the platform.
See also: NVIDIA T3000 and T2000 target robotics cost and power limits
Want to learn more about the IoT from industry leaders? Check out IoT Tech Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including AI & Big Data Expo and the Cyber Security Expo. Click here for more information.
IoT News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
