NVIDIA Opens Alpamayo 2 Edge for Onboard AI Driving

By Grace Kim

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

What does the Edge release change?

NVIDIA first announced Alpamayo 2 Edge in its October 21, 2026 Hyperion announcement. The company said Edge was optimized for DRIVE AGX Thor and designed for 10 Hz in-vehicle operation. The Alpamayo product page lists Edge as available and separates it from the 34-billion-parameter Alpamayo 2 Super. NVIDIA says Super is intended for driving prediction, scene understanding, auto-labeling and training smaller models. The product page thus puts a compact model on the vehicle and a larger model upstream in the training workflow.

How does Edge reason and produce driving actions?

NVIDIA says Cosmos 3 Edge supplies the Nemotron-based reasoning backbone, while a diffusion action expert produces driving actions. In the vision-language-action design, surround-camera video, temporal context, ego-motion history and navigation text feed future trajectories and Chain-of-Causation explanations linking decisions to causal factors. Edge can also answer free-form questions about a driving scene. For routine driving, Edge can omit a reasoning trace instead of producing one for every decision. In those cases, NVIDIA says the inference script emits special NOMINAL tokens, reducing token generation and inference compute while the model continues to produce trajectories. The model can explain a decision when needed and stay lean when it does not.

What data and training shaped Edge?

The model card lists a global deployment geography, OpenMDW-1.1 licensing for the weights and Apache License 2.0 for the source code. Training data comprise roughly 115,000 hours of multi-camera driving video with ego-motion and trajectory annotations, plus about 3.7 million Chain-of-Causation reasoning traces. NVIDIA says the dataset includes camera, inertial measurement unit and GPS records, as well as synthetic reasoning traces. For training, Alpamayo 2 Super generated the reasoning autolabels used for offline distillation. Edge then learned from the consequences of its actions through closed-loop reinforcement learning in the AlpaGym simulator. NVIDIA reports that this stage improved closed-loop driving performance by 48.9% over the best checkpoint without closed-loop reinforcement learning. Super supplied the labels, while reinforcement learning connected those labels to closed-loop behavior.

Can Edge meet onboard latency requirements?

The model card says the downloaded BF16 weights run directly as a PyTorch model on supported NVIDIA GPUs, but real-time operation on DRIVE AGX Thor requires further optimization. NVIDIA’s deployment recipe applies FP8 calibration, ONNX export, TensorRT and TensorRT-Edge-LLM, with DFlash2 speculative decoding for online reasoning. The recipe reports 160–190 milliseconds between camera input and trajectory output, depending on the amount of reasoning. The measurement used DRIVE AGX Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4 and TensorRT 11.3.1.9. In AlpaSim, NVIDIA reports that Edge traveled 9.87 kilometers between at-fault incidents, versus 0.25 kilometers for the 4-billion-parameter Qwen-Drive-1.0. The same comparison produced minimum average displacement error of 0.53 meters for Edge and 1.68 meters for Qwen-Drive-1.0. NVIDIA notes that lower values indicate closer trajectory predictions. The model card lists baseline software testing on one NVIDIA H100 80GB GPU and says other GPU architectures have not yet been validated. That baseline statement is separate from the optimized DRIVE AGX Thor-X recipe. Edge now has a documented deployment path, but its real-time credentials still depend on platform-specific optimization.

How does AlpaJudge connect evaluation to training?

NVIDIA also released AlpaJudge, a human-calibrated evaluation and reward model covered by the same weight and code licenses.

• AlpaJudge-RB takes a question, reference answer and candidate answer without images. It returns one alignment probability from 0 to 1.
• AlpaJudge-RF uses three front-wide and three front-tele frames, ego meta-action history and future context. It also takes a candidate rationale and returns five rubric probabilities, plus action, component and overall scores.

NVIDIA’s guidance says a high reasoning score provides evidence that a rationale matches AlpaJudge’s rubric, but collision, road-departure, close-encounter and trajectory metrics remain necessary to describe driving behavior. The company says low-scoring examples deserve review rather than automatic rejection because they can expose rare failures. Very high scores can identify scenes the model already explains consistently, while middle scores often mark useful further-training opportunities. The public reinforcement-learning recipe can use AlpaJudge-RB to grade sampled reasoning together with trajectory error and comfort. NVIDIA recommends retaining the trajectory gate and logging component scores so a fluent but incorrect explanation cannot alone earn a high total reward. AlpaJudge turns explanation quality into a measurable input for the next training cycle.

What remains before deployment?

The model card says integration into an autonomous-driving system requires additional testing with use-case-specific data and iterative unit- and system-level validation before deployment. Developers can now test compact onboard inference, inspect scored explanations and retrain against simulator failures. The next test is whether those gains survive use-case-specific validation and translate into safer real-vehicle behavior.