Source: Unite.AI
NVIDIA released the Alpamayo 2 Edge model weights on Hugging Face on November 10, 2026, making a 2-billion-parameter model available to autonomous-driving developers working within 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, describing a model optimized for DRIVE AGX Thor and designed for 10 Hz in-vehicle operation. The Alpamayo product page lists Edge as available and separates the compact in-vehicle model from the 34-billion-parameter Alpamayo 2 Super, which is intended for driving prediction, scene understanding, auto-labeling and training smaller models.
Compact Model With Adaptive Reasoning
Cosmos 3 Edge supplies the Nemotron-based reasoning backbone, while a diffusion action expert produces driving actions. In this vision-language-action design, surround-camera video, temporal context, ego-motion history and navigation text are converted into future trajectories and Chain-of-Causation explanations linking decisions to causal factors. The model can also answer free-form questions about a driving scene.
Edge can omit a reasoning trace for routine driving instead of producing one for every decision. In those cases, the inference script emits special NOMINAL tokens, cutting token generation and inference compute while the model continues to produce trajectories.
The model card lists a global deployment geography, OpenMDW-1.1 licensing for the weights and Apache License 2.0 for the source code. Its training dataset comprises 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. The data include 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 subsequently 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.
Onboard Latency and Validation Limits
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 recipe uses 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, measured on DRIVE AGX Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4 and TensorRT 11.3.1.9.
NVIDIA reports that Edge recorded 9.87 kilometers between at-fault incidents in the AlpaSim closed-loop simulator, versus 0.25 kilometers for the 4-billion-parameter Qwen-Drive-1.0. The same comparison put minimum average displacement error at 0.53 meters for Edge and 1.68 meters for Qwen-Drive-1.0, with lower values indicating closer trajectory predictions.
For baseline software compatibility, the model card lists testing on one NVIDIA H100 80GB GPU and says other GPU architectures have not yet been validated, a separate statement from the optimized Thor-X recipe. Integration into an autonomous-driving system requires additional testing with use-case-specific data and iterative unit- and system-level validation before deployment.
AlpaJudge Connects Evaluation to Training
NVIDIA also released AlpaJudge, a human-calibrated evaluation and reward model with the same weight and code licenses. AlpaJudge-RB is reference-based: it takes a question, reference answer and candidate answer without images and returns one alignment probability from 0 to 1. The reference-free AlpaJudge-RF instead uses three front-wide and three front-tele frames, ego meta-action history and future context, and a candidate rationale to return five rubric probabilities plus action, component and overall scores.
NVIDIA’s guidance treats a high reasoning score as evidence that a rationale matches the judge’s rubric, with collision, road-departure, close-encounter and trajectory metrics still needed to describe driving behavior. 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 further-training opportunities.
The public reinforcement-learning recipe can use AlpaJudge-RB to grade sampled reasoning together with 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. The release’s practical 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.
