NVIDIA Alpamayo 2 Edge is a compact, open-weights AI model that reads car-camera video, plans a driving path and explains why, and NVIDIA released it on Hugging Face on 10 November 2026. It has about 2 billion parameters and is designed to run inside a car on NVIDIA’s DRIVE AGX Thor computer. This guide explains what Alpamayo 2 Edge is, what NVIDIA has proven, what it has not, and what an Indian developer can actually do with it today.
Key takeaways
- What: a 2B-parameter vision-language-action model that outputs a driving trajectory plus a short causal explanation, built on NVIDIA Cosmos 3 Edge.
- When: announced on 21 October 2026 at GTC Berlin; weights and code released on 10 November 2026.
- Speed: NVIDIA reports 160 to 190 ms from camera pixels to trajectory on DRIVE AGX Thor after optimisation.
- Licence: weights under OpenMDW-1.1, code under Apache 2.0.
- Catch: the headline safety numbers come from NVIDIA’s own simulator, and NVIDIA says the model needs more testing before any vehicle deployment.
Published: 11 November 2026. Last updated: 11 November 2026, IST. Status: released on 10 November; independent road tests are not yet available.
Table of Contents
Table of Contents
What is NVIDIA Alpamayo 2 Edge?
Alpamayo 2 Edge is a vision-language-action (VLA) model: it takes in pictures and text, and outputs actions, here a driving path. According to NVIDIA’s Hugging Face model card, it uses surround-camera video, navigation instructions and the car’s recent movement to produce a future trajectory and a Chain-of-Causation explanation, a short chain linking what it sees to what it decides.
Think of a driver who talks only when something unusual happens. On an empty road the model stays quiet and emits a special NOMINAL token, and near a pedestrian or a stopped bus it explains its reasoning. NVIDIA calls this adaptive reasoning, and it saves compute inside a car where every millisecond counts.
What is the Alpamayo 2 Edge release date?
The Alpamayo 2 Edge release date was 10 November 2026, when NVIDIA published the weights on Hugging Face and a developer blog. NVIDIA first announced it on 21 October at GTC Berlin and said the download would come later in October, so the release landed about ten days after that window by my reading.
| Date | What happened | Source |
|---|---|---|
| 5 January 2026 | NVIDIA announces the Alpamayo family of open models, simulation tools and datasets | Yahoo Finance |
| June to August 2026 | Alpamayo 2 Super, a much larger model for Level 4 robotaxis, is reported and released (32B to 34B parameters, outlets differ) | AI Weekly, daily.dev |
| 21 October 2026 | Alpamayo 2 Edge announced at GTC Berlin with Hyperion partners | NVIDIA |
| 5 November 2026 | NVIDIA staff on the developer forum say the weights are about a week away | NVIDIA forum |
| 10 November 2026 | Weights, code and developer blog go live | Hugging Face, NVIDIA |
How does the model work?
The model pairs a small language-and-vision backbone with a diffusion-based action expert. The backbone reads the scene and writes the reasoning. The action expert then generates the driving path. NVIDIA’s technical blog lists the main ingredients below.
- Distillation: the larger Alpamayo 2 Super produced reasoning labels that Edge learned from.
- Closed-loop reinforcement learning: in NVIDIA’s AlpaGym simulator the model learns from the results of its own actions. NVIDIA says this improved driving performance by 48.9% over the best checkpoint without it.
- Speculative decoding (DFlash2): a small drafter guesses ahead so the main model writes reasoning faster.
- Multi-resolution inputs: cheaper encoding of the surround-view cameras.
- TensorRT recipes: instructions to compress and compile the model for the car’s chip.
What are the Alpamayo 2 Edge specs?
| Item | Detail |
|---|---|
| Parameters | Headline “2B”. Model card: 2.45B backbone plus 0.2B action expert; Hugging Face lists 3B |
| Base model | NVIDIA Cosmos 3 Edge (Nemotron-based backbone) with a diffusion action decoder |
| Inputs | Six cameras, 22 frames, navigation text, 20 past ego poses at 0.1-second steps |
| Outputs | Chain-of-Causation text and 64 waypoints covering 0.1 to 6.4 seconds ahead |
| Weights licence | OpenMDW-1.1; source code under Apache 2.0 |
| Training data | About 115,000 hours of multi-camera driving video and about 3.7 million reasoning traces |
| Tested hardware | One NVIDIA H100 80GB; peak 12.06 GiB memory; NVIDIA recommends at least 16 GB |
| Target hardware | NVIDIA DRIVE AGX Thor, designed for 10 Hz operation |
Parameter count is a source conflict. Adding the model card’s two figures gives about 2.65 billion, my sum, and the Hugging Face page shows 3B. NVIDIA’s “2B” looks like a rounded label for the language backbone, not the full model.
How fast is Edge on DRIVE AGX Thor?
After optimisation, Alpamayo 2 Edge runs with a pixel-to-trajectory latency of 160 to 190 ms on DRIVE AGX Thor-X, depending on how much reasoning it writes, according to NVIDIA. The recipe uses FP8 calibration, ONNX export, TensorRT and DFlash2, and it was measured with DRIVE OS 7.2.7.0, CUDA 13.4 and TensorRT 11.3.1.9.
Partner Embedl says it had the model planning at 10 Hz on AGX Thor in under 48 hours. One open question is mine: 10 Hz means a new plan every 100 ms, while 160 to 190 ms is end-to-end delay. NVIDIA’s blog does not explain how those two numbers fit, so ask before building on either.
The hardware limits are clear. NVIDIA staff said on the forum that the older Jetson AGX Orin “will be too slow to run at production speeds”, and the model card validates only the H100. The default BF16 weights run in PyTorch on supported NVIDIA GPUs, but real-time use needs the Thor recipe.
How good are the Alpamayo 2 Edge benchmarks?
On NVIDIA’s own tests, Alpamayo 2 Edge drives about 15 km between at-fault incidents in simulation and has best-of-six predicted paths that average about half a metre from recorded driving. These Alpamayo 2 Edge benchmarks were run by NVIDIA, on its simulator and datasets, so they are claims until others repeat them.
| Test | Alpamayo 2 Edge result |
|---|---|
| AlpaSim closed-loop (875 scenarios, 3 runs each) | 15.15 km per at-fault incident (off-road or at-fault collision) |
| Open-loop minADE at 5 s (1,392 samples) | 0.52 m (lower is better) |
| LingoQA, Lingo-Judge score | 75.2 |
| Ego3D accuracy | 84.94 |
| SURDS | 86.50 |
| WaymoQA, all / safety-critical | 75.77 / 71.09 |
| nuReasoning, choice accuracy | 78.58 |
| KITScenes, scene | 89.46 |
NVIDIA also compares the model with Qwen-Drive-1.0, a 4B-parameter rival, and says Edge goes about 60 times further between at-fault incidents. Unite.AI reports Qwen-Drive-1.0 at 0.25 km, and 15.15 divided by 0.25 is about 61, my arithmetic, which fits NVIDIA’s chart.
Source conflict: Unite.AI, an AI-written site, reports 9.87 km for Edge, and a Reddit post says “40 times”. Both fit each other but not the model card. I use the model card because it is NVIDIA’s current primary document, but the gap shows the numbers moved between versions.
Partners supplied early results, which NVIDIA published. PORT says that in simulation on its London street data, Edge kept its lane in more than twice as many runs as Alpamayo 1.5. Niulinx says it cut displacement error by 13.6% against Alpamayo 2 Super in a zero-shot test and by 44.2% more after fine-tuning. Treat these as vendor-selected quotes.
Edge vs Alpamayo 2 Super: what is the difference?
Alpamayo 2 Edge is the small model that runs in the car, and Alpamayo 2 Super is the large model that works in the data centre. Super teaches Edge: it writes reasoning labels that Edge learns from. NVIDIA’s press release says Edge is designed for in-vehicle scene understanding and motion planning.
| Feature | Alpamayo 2 Edge | Alpamayo 2 Super |
|---|---|---|
| Size | About 2B (headline) | 32B to 34B, outlets differ |
| Main job | Run inside the vehicle on DRIVE AGX Thor | Prediction, scene understanding, auto-labelling, training smaller models |
| Released | 10 November 2026 | Mid-2026 |
| Weights licence | OpenMDW-1.1 | Reported as commercially usable |
| Needs | 16 GB or more GPU memory to test | Far larger GPU setup |
What is AlpaJudge and why does it matter?
AlpaJudge is a separate NVIDIA model that scores how good a driving explanation is. It comes in two forms: a reference-based version that compares an explanation with an expert’s, and a reference-free version that checks it against camera frames and the car’s actions. NVIDIA used it to pick training data and as a reward during reinforcement learning.
NVIDIA also warns that a high reasoning score is only evidence that an explanation matches a rubric, so it should be reported next to collision and trajectory results. That caution is important because a fluent explanation can still describe a wrong decision.
Who is building with NVIDIA’s Alpamayo platform?
NVIDIA’s 21 October announcement named Volvo Cars, Mercedes-Benz, Bolt with Lucid, Volkswagen Commercial Vehicles and Einride among partners on the Hyperion platform. Bolt plans to start gathering driving data in Madrid in 2027, subject to regulatory approval, as part of a planned European rollout of at least 25,000 Level 4-capable Lucid vehicles.
These partners adopt the wider Alpamayo Stack and Hyperion hardware. NVIDIA’s release does not say any of them has put Edge itself into a car. The early Edge testers it names are PORT, Niulinx and Embedl, all of which are testing, not selling, a product.
Physical AI is spreading beyond cars. Anthropic’s new Fable 5.2 also reports robot-control results, covered in our Claude Fable 5.2 guide, and we have tracked the Meta humanoid robot effort.
What are the risks and limits?
- Simulation is not the road. The 15.15 km figure comes from AlpaSim, NVIDIA’s simulator. It is not a measure of real-world safety.
- Narrow validation. The model card tested only an H100 GPU and says other GPUs are unvalidated. The Thor recipe is a separate check.
- Not a finished product. NVIDIA says integrating a foundation model needs use-case-specific testing and unit- and system-level validation before deployment.
- Explanations may mislead. A neat causal story is not proof the model decided for that reason.
- Vendor-picked evidence. Partner quotes and benchmark charts come from NVIDIA’s own publication.
Safety evidence is a bigger topic. NVIDIA introduced a framework called Halos SAFE on 21 October to produce traceable safety evidence from simulation and road tests. Our report on the humanoid robot safety conference covers the same question for robots.
Alpamayo 2 Edge in India: what can developers do?
The weights are listed with a global deployment geography, so Indian developers, students and startups can download Alpamayo 2 Edge from Hugging Face like anyone else. A first test needs a Linux machine with an NVIDIA GPU of 16 GB or more, which NVIDIA recommends but has not validated beyond the H100.
Real-time car use needs DRIVE AGX Thor, which is automaker-grade hardware, so most Indian users will work in simulation or on recorded data. NVIDIA’s model card does not list Indian roads in its training data, and Indian traffic is mixed, dense and less regular. Treat out-of-the-box results on Indian video as unproven.
There is still a real opening: open weights let Indian engineering teams study reasoning-based driving, fine-tune on local clips and benchmark it against their own stacks. Niulinx reported a 44.2% error cut by fine-tuning on its own data, which suggests local data matters. Chip supply and costs shape all of this, as in our chip stocks selloff guide and the Microsoft Maia 300 chip report.
How to try the model step by step
- Use Linux with Python 3.12, PyTorch 2.8.0, torchvision 0.23.0 and Transformers 5.17.0, as listed on the model card.
- Make sure you have an NVIDIA GPU with at least 16 GB of memory. FlashAttention 2.8.3 is optional but speeds up inference.
- Open the Alpamayo 2 Edge Hugging Face page, create an account, log in with
hf auth loginand accept the PhysicalAI Autonomous Vehicles dataset terms for the examples. - Clone the code repository, load
nvidia/Alpamayo2-Edgeand run the inference script on a sample clip. - Open
output.json, read the reasoning undertext.cotand the trajectory, and note anyNOMINALtokens. - Use the chat template with
enable_thinking=False; the model card says the template turns thinking on by default. - Only after that try AlpaJudge, which needs roughly 5 GB for weights, to score the explanations.
Common mistakes and myths
- Myth: it is a self-driving car you can buy. It is a model for developers, with no road approval attached.
- Myth: it is exactly 2.0 billion parameters. The model card’s parts add up to about 2.65 billion.
- Mistake: running it on a Jetson Orin and expecting real-time results. NVIDIA staff recommend Thor.
- Mistake: treating the 60 times figure as road safety. It is a simulator comparison against one rival.
- Myth: open weights mean no rules. Read the OpenMDW-1.1 licence and the dataset terms before building a product.
What happens next?
- Agentic Factory: NVIDIA says its Alpamayo Agentic Factory blueprint is expected in early 2027.
- More numerics: NVIDIA staff said the quantisation recipe will likely be updated later to support other precisions.
- Europe: Mercedes-Benz said its MB.DRIVE ASSIST PRO driver-assistance system will soon get approval in Germany; Bolt’s Madrid data gathering is planned for 2027.
- Independent tests: outside results on Edge should appear in the coming weeks.
- This page: Askhari will update this same URL with dated notes.
Related reading on Askhari
- Claude Fable 5.2 and Mythos 5.2 explained
- Humanoid robot safety at Detroit’s robot safety conference
- Meta humanoid robot Max explained
- Chip stocks selloff: why Intel sent AI chips down
- Microsoft Maia 300 AI chip
What we don’t know yet
- How Alpamayo 2 Edge performs on real roads, and on Indian roads.
- Whether the 160 to 190 ms latency fits the 10 Hz planning loop in a production stack.
- Which automakers, if any, will put Edge into production vehicles.
- Why published safety numbers differ between NVIDIA’s card and secondary reports.
- Whether independent testers can reproduce NVIDIA’s closed-loop results.
Watch: NVIDIA Alpamayo 2 Super deep dive
Frequently asked questions about Alpamayo 2 Edge
What is Alpamayo 2 Edge?
It is a roughly 2-billion-parameter NVIDIA model that turns camera video and navigation text into a driving trajectory and a causal explanation, designed for the DRIVE AGX Thor computer.
Is it open source?
The weights are released under the OpenMDW-1.1 licence and the code under Apache 2.0. That is open weights with licence terms, so read the licence before commercial use.
Can it drive a car today?
Not by itself. NVIDIA says the model is a starting point that needs extra testing and validation before it goes into a vehicle.
What hardware does it need?
For real-time use, NVIDIA points to DRIVE AGX Thor. For experiments, an NVIDIA GPU with at least 16 GB of memory is recommended, though only the H100 is validated.
What is the difference between Edge and Super?
Edge is the small in-car model, and Super is a much larger data-centre model that helps train it by writing reasoning labels.
Can I use Alpamayo 2 Edge in India?
Yes, you can download it from Hugging Face. Expect to fine-tune on local data, since NVIDIA does not say it was trained on Indian roads.
Hari’s take
Alpamayo 2 Edge is the first time I have seen a reasoning driving model small enough to talk about running in a car, with open weights anyone can download. That is a big deal for students and startups. But a simulator score is not a road test, and NVIDIA’s own numbers shifted between versions.
Use it to learn and to prototype, not to make safety claims. If you know a friend working on ADAS, robotics or autonomous-vehicle research, send them this guide, since they will want the hardware limits and licence details in one place.
Sources: NVIDIA Technical Blog, 10 Nov 2026 · Hugging Face model card · NVIDIA press release, 21 Oct 2026 · NVIDIA blog, 21 Oct 2026 · NVIDIA developer forum, 5 Nov 2026 · Unite.AI, 10 Nov 2026 · Yahoo Finance, 5 Jan 2026 · AI Weekly · daily.dev · Reddit r/aicuriosity. Last updated 11 November 2026.
