NVIDIA adds Alpamayo 2 Edge, a 2B open self-driving model for in-car use

NVIDIA’s Alpamayo product page now lists Alpamayo 2 Edge, a 2-billion-parameter vision-language-action model it describes as built “for in-car reasoning”, alongside the 34-billion-parameter Alpamayo 2 Super. The most important fact is the size: Edge is a fraction of Super’s scale and smaller than the 10B Alpamayo 1 Nano and 1.5 Nano models that came before it.

That one line is, for now, nearly all the public detail on Edge. The sources contain no model card, benchmark, camera configuration or hardware target for it. What they do describe in depth is the family Edge joins, the licence NVIDIA applies to that family, and the larger model it sits next to.

What NVIDIA says about Alpamayo 2 Edge

On its Alpamayo page, NVIDIA groups its “Open VLA Models” into two entries: “Alpamayo 2 Edge: 2B VLA for in-car reasoning” and “Alpamayo 2 Super: 34B reasoning foundation model for planning, auto-labeling, and evaluation”, followed by a pointer to download on Hugging Face. A vision-language-action (VLA) model takes in camera images and text and produces both language and physical actions, in this case a driving path.

The split in wording matters. Super is described as a development tool for planning, auto-labelling (machine-generated training annotations) and evaluation, while Edge is described by where it runs: in the car. The same page presents an “Alpamayo Stack”, full-stack driving software on NVIDIA Hyperion that includes “an in-vehicle reasoning model that interprets road scenes and generates driving trajectories”. The page does not say whether that in-vehicle model is Edge.

NVIDIA describes the family’s distinguishing feature as Chain of Causation reasoning: the model explains why it chose a driving action, with the aim of handling rare scenarios and making decisions explainable.

How Edge fits in the Alpamayo family

NVIDIA first announced Alpamayo on January 5, 2026 at CES 2026. Alpamayo 2 Super was announced on May 31, 2026 at GTC Taipei, with weights released on August 4, 2026. Combining sizes from NVIDIA’s pages gives this picture:

Model	Parameters	Stated role
Alpamayo 2 Edge	2B	In-car reasoning
Alpamayo 2 Super	34B	Planning, auto-labelling and evaluation
Alpamayo 1.5 Nano	10B (8B + 2B)	Reinforcement-learning post-trained; adds navigation guidance and text-steered planning
Alpamayo 1 Nano	10B (8B + 2B)	Trajectory plus causal trace from four-camera video

The two NVIDIA pages describe the family differently. NVIDIA’s developer reference page, which says it was last updated in August 2026, lists three models (1 Nano, 1.5 Nano and 2 Super) and states that “Alpamayo models are cloud-side teacher models, not in-vehicle runtimes”. That page says developers can fine-tune and distil (compress a large model’s behaviour into a smaller one) the models into compact student models that run on DRIVE AGX Thor. The product page instead lists Edge and Super only, with Edge labelled for in-car reasoning. Builders reading NVIDIA’s documentation should expect these two descriptions to disagree until NVIDIA reconciles them.

NVIDIA’s August announcement for Super described a “cloud-to-car workflow”: frontier-scale reasoning in the cloud, with distilled models “optimized for efficient, real-time inference in production vehicles”. A 2B model labelled for in-car use fits the vehicle side of that description, though NVIDIA’s page does not say how Edge was trained or whether it was distilled from Super.

What Alpamayo 2 Super brings, in NVIDIA’s own numbers

Because Edge has no published results, Super is the only Alpamayo 2 model with detailed figures. Super combines a 32-billion-parameter Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based Action Expert and is post-trained with reinforcement learning. NVIDIA’s reference page lists the action expert at 2.3B, a small discrepancy between NVIDIA’s own documents.

Super processes up to seven cameras for 360-degree coverage and produces trajectories, Chain-of-Causation traces, meta-actions such as yield or stop, answers to questions about the scene with 2D grounding, and structured auto-labels. Its trajectory output is 64 waypoints. NVIDIA says Super was trained on roughly 115,000 hours of multi-camera driving video, more than 1 billion images and roughly 3,700,000 reasoning traces, a corpus it holds privately.

All benchmark figures below come from NVIDIA’s own evaluation:

• On LingoQA, a driving reasoning benchmark, Super scores 79.2, leading Qwen2.5-VL 72B by 17.0 points.
• In NVIDIA testing with the Lingo-Judge metric, it outperformed Gemini 2.5 Pro by 15.1 points and GPT-4o by 23.2 points, and NVIDIA says it ranks first among nearly 40 models evaluated.
• Open-loop evaluation on 1,434 challenging samples gives a 6.4-second minADE_6 (average distance between predicted and actual path, best of six tries) of 0.911 m.
• Closed-loop simulation in NVIDIA AlpaSim on 913 reconstructed scenes yields an AlpaSim Score of 1.50 ± 0.13, which NVIDIA says reveals collisions and close encounters that open-loop metrics miss.

None of these numbers describe Edge, and no independent benchmark appears in the sources for either model.

Licence and access

NVIDIA’s product page places Edge among its open VLA models and directs developers to Hugging Face for downloads. The sources do not show a separate Hugging Face repository for Edge, so its exact repository name and files are not confirmed here.

For the family as a whole, NVIDIA said in August that the OpenMDW licence “is now being applied across the entire Alpamayo model family so developers can deploy any of the models commercially without requiring additional permissions”. OpenMDW-1.1 is the Linux Foundation’s permissive licence for open model distributions and covers fine-tuning, derivative models and commercial redistribution. NVIDIA adds that distilled models can be deployed commercially without further permission and that model outputs carry no licence conditions. Code is released under Apache-2.0. The sources do not state Edge’s licence by name, so developers should confirm it on the model’s own listing.

Super’s Hugging Face repository is public rather than gated, tagged under the OpenMDW-1.1 licence and built on Cosmos3-Super, with 70,732 all-time downloads at the time the page was captured. NVIDIA reports more than 600,000 Hugging Face downloads across Alpamayo as of August 2026. For hosted inference, NVIDIA lists Alpamayo R1.5 as a NIM endpoint; no hosted option for Edge is mentioned.

The tooling around the models

NVIDIA’s open platform pairs the models with AlpaSim for closed-loop evaluation of driving decisions and AlpaGym for closed-loop reinforcement learning. Its Physical AI Open Datasets cover driving scenarios across 25 countries and include Chain-of-Causation reasoning labels. NVIDIA says the platform ships training recipes for fine-tuning on proprietary driving data.

On the vehicle side, the Alpamayo Stack runs on NVIDIA Hyperion and adds runtime layers around the reasoning model: Halos SAFE for simulation-based validation, a Policy Evaluator and Safety Monitors that evaluate driving actions and support trajectory selection, and Halos OS for system services with classical perception and planning. That layering matters for anyone picturing a small model alone driving a car: NVIDIA’s design surrounds the in-vehicle model with independent safeguards.

What it means for builders, and the open questions

For teams working on autonomous driving or other physical AI, a 2B open driving model is a different proposition from a 34B one. NVIDIA positions Super for cloud work such as generating reasoning traces, synthetic training data and teacher outputs for distillation. A model sized and labelled for the car is the piece that would sit on the production side of that workflow, if it performs well enough.

Several questions remain open in the sources:

• Performance. There are no Edge results on LingoQA, AlpaSim or any other benchmark, and no comparison with Super or the 10B Nano models.
• Inputs and outputs. It is not stated whether Edge keeps Super’s seven-camera input, its Chain-of-Causation traces or its question-answering outputs.
• Hardware. NVIDIA’s reference page names DRIVE AGX Thor as the target for distilled student models, but no source ties Edge to a specific chip, latency or memory budget.
• Validation. Foretellix, which works with the Alpamayo ecosystem, argues that as driving models become more capable and widely available, validating them becomes harder, and that developers must show where a model has and has not been tested. It notes that UNECE WP.29 R185, taking effect in January 2027, requires manufacturers to demonstrate coverage of the operational design domain in their safety case.

For context on open models aimed at physical systems, see Unicorn’s coverage of Runway’s Praxis-1 robot-control model and DeepMind’s Gemini Robotics 2.5 and its robot safety benchmark.

What to watch next

The first thing to look for is an Edge model card: parameter breakdown, camera inputs, outputs, training method and licence. The second is any benchmark, ideally on the same LingoQA and AlpaSim tests NVIDIA used for Super, so Edge can be compared with its larger sibling. The third is whether NVIDIA’s developer reference page, which still describes Alpamayo models as cloud-side teachers, is updated to cover an in-car model. Finally, with R185 arriving in January 2027, watch for validation work from partners such as Foretellix on how small in-vehicle models like Edge are tested before deployment.