NVIDIA releases Alpamayo 2 Edge weights for autonomous driving

NVIDIA released the 2-billion-parameter Alpamayo 2 Edge model weights on Hugging Face on November 10, 2026, giving autonomous-driving developers a compact model built for onboard compute and latency limits. The release includes deployment recipes and evaluation tools for producing and scrutinizing driving trajectories and their explanations.

NVIDIA first announced Alpamayo 2 Edge in its October 21, 2026, Hyperion announcement. NVIDIA described the model as optimized for DRIVE AGX Thor and designed for 10 Hz in-vehicle operation.

The stakes are measurable. NVIDIA reports 160–190 milliseconds from camera input to trajectory output on DRIVE AGX Thor-X, plus a 48.9% improvement in closed-loop driving performance after reinforcement learning in the AlpaGym simulator. According to NVIDIA, the package links compact onboard inference with scored explanations and targeted retraining, creating a development loop for testing whether better explanations produce better driving.

What did NVIDIA release, and how does it fit the larger model?

The Alpamayo product page lists Edge as available and separates it from the 34-billion-parameter Alpamayo 2 Super. NVIDIA intends Super for driving prediction, scene understanding, auto-labeling and training smaller models.

Super also generated the reasoning autolabels, or machine-generated training labels, used to train Edge through offline distillation. Distillation transfers those labels from the larger model to the compact one.

The smaller model combines two main components. Cosmos 3 Edge supplies the Nemotron-based reasoning backbone, while a diffusion action expert, a generative module for driving actions, produces the vehicle's actions.

In this vision-language-action design, a system combining video perception, language reasoning and action output, Edge accepts:

• Surround-camera video.
• Temporal context.
• Ego-motion history, or records of the vehicle's own movement.
• Navigation text.

NVIDIA says Edge converts those inputs into future trajectories and Chain-of-Causation explanations. Those explanations link driving decisions to causal factors. The model can also answer free-form questions about a driving scene.

Why does Edge skip some reasoning traces?

NVIDIA says: "Edge can omit a reasoning trace for routine driving instead of producing one for every decision."

For routine maneuvers, the inference script emits special NOMINAL tokens. These tokens reduce token generation and inference compute while Edge continues producing trajectories. The amount of reasoning also affects latency, which explains NVIDIA's 160–190 millisecond range rather than a single fixed figure.

That conditional design gives developers a way to reserve deeper reasoning for situations that require it. NVIDIA provides both the model's reasoning trace and the option to suppress it during routine operation.

What data and licenses did NVIDIA publish?

The model card lists a global deployment geography. It assigns OpenMDW-1.1 licensing to the weights and Apache License 2.0 to the source code. The model card splits licensing two ways.

NVIDIA reports a training dataset containing roughly 115,000 hours of multi-camera driving video with ego-motion and trajectory annotations. The dataset also includes about 3.7 million Chain-of-Causation reasoning traces.

The records cover three main categories:

• Camera data.
• Inertial measurement unit data, covering motion and orientation.
• GPS records.

NVIDIA also includes synthetic reasoning traces alongside the recorded driving data. Developers must account for both the stated dataset scale and the separate licensing terms for weights and code.

How much did reinforcement learning improve Edge?

Alpamayo 2 Super generated the reasoning autolabels used for offline distillation. Edge then learned from the consequences of its own actions through closed-loop reinforcement learning, which trains a model against feedback from its actions in a simulator, inside AlpaGym.

NVIDIA reports that this reinforcement-learning stage improved closed-loop driving performance by 48.9% over its best checkpoint without closed-loop reinforcement learning. The 48.9% gain is simulator-based.

The training sequence therefore combines two distinct stages. Offline distillation transfers reasoning supervision from the 34-billion-parameter Super model, while reinforcement learning exposes Edge to action consequences inside AlpaGym.

Can Edge meet real-time latency inside a vehicle?

The downloaded BF16 weights run directly as a PyTorch model on supported NVIDIA GPUs. BF16 is a 16-bit floating-point format. NVIDIA says real-time operation on DRIVE AGX Thor still requires further optimization.

Its deployment recipe uses:

• FP8 calibration, an 8-bit floating-point conversion step.
• ONNX export, a model interchange step.
• TensorRT and TensorRT-Edge-LLM inference software.
• DFlash2 speculative decoding, which drafts multiple tokens at once to accelerate online reasoning.

The recipe reports 160–190 milliseconds between camera input and trajectory output. The range depends on how much reasoning Edge performs.

NVIDIA measured those figures on DRIVE AGX Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4 and TensorRT 11.3.1.9. Those results belong to that specific optimized configuration.

The model card separately lists baseline testing on one NVIDIA H100 80GB GPU. It says other GPU architectures have not yet been validated, a statement distinct from the optimized Thor-X recipe.

NVIDIA still requires deployment validation. Integration demands additional testing with use-case-specific data and iterative unit- and system-level validation before deployment.

How did Edge compare with Qwen-Drive-1.0?

According to NVIDIA, Edge recorded 9.87 kilometers between at-fault incidents in the AlpaSim closed-loop simulator. The 4-billion-parameter Qwen-Drive-1.0 recorded 0.25 kilometers in the same comparison.

The comparison also produced these trajectory measurements:

• Minimum average displacement error of 0.53 meters for Edge.
• Minimum average displacement error of 1.68 meters for Qwen-Drive-1.0.

NVIDIA says lower displacement-error values indicate closer trajectory predictions. Both comparisons come from AlpaSim, and NVIDIA's release reports no public-road results for either model.

What does AlpaJudge measure, and how does it feed training?

NVIDIA also released AlpaJudge, a human-calibrated evaluation and reward model with the same weight and code licenses as Edge. A reward model scores outputs so those scores can serve as a training signal.

AlpaJudge comes in two forms. AlpaJudge-RB is reference-based, taking a question, a reference answer and a candidate answer without images. It returns one alignment probability ranging from 0 to 1.

AlpaJudge-RF is reference-free. It uses six camera frames, consisting of three front-wide and three front-tele images, alongside ego meta-action history, future context and a candidate rationale. The wide-angle and telephoto views cover the vehicle's forward field of view.

The reference-free model returns five rubric probabilities, with each probability representing conformity to one scoring criterion. It also returns action, component and overall scores.

NVIDIA treats a high reasoning score as evidence that a rationale matches the judge's rubric. Collision, road-departure, close-encounter and trajectory metrics remain necessary to describe driving behavior.

The guidance says: "Low-scoring examples should be reviewed rather than automatically discarded because they can expose rare failures." Very high scores may identify scenes the model already explains consistently, while middle scores often mark useful opportunities for further training.

The public reinforcement-learning recipe can use AlpaJudge-RB to grade sampled reasoning alongside trajectory error and comfort. NVIDIA recommends keeping the trajectory gate and logging component scores so a fluent but incorrect explanation cannot alone produce a high total reward.

AlpaJudge scores explanations, not driving safety. NVIDIA's safeguards keep trajectory evidence inside the reward calculation, which makes the evaluation output usable for targeted retraining rather than simple rejection.

What changes for autonomous-driving developers?

According to NVIDIA, the release's significance is the development loop it enables: compact onboard inference, scored explanations and targeted retraining, with collision and trajectory evidence determining whether better explanations translate into better driving.

For adopters, the next test is their own use-case data, target hardware and validation process. NVIDIA's reported latency comes from a specific Thor-X configuration, while its baseline model-card testing covers one H100 80GB GPU. Deployment readiness therefore depends on reproducing the measured behavior within each developer's own system.