Nvidia Releases Alpamayo 2 Edge, a 2-Billion-Parameter Autonomous Driving Model That Explains Its Choices Only When Needed

Nvidia released Alpamayo 2 Edge on November 10, 2026, a vision-language-action autonomous driving model of approximately 2 billion parameters, designed to run directly inside the vehicle on the DRIVE AGX Thor chip. It processes inputs from six cameras, plans trajectories for the next 6.4 seconds, and, whenever the situation justifies it, writes out an explanation for its decision. The weights are available for download on Hugging Face under the OpenMDW-1.1 license—a permissive license that does not restrict commercial use—and the code is released under Apache 2.0.

Explaining Only When Necessary

“Reasoning” driving models suffer from an inherent drawback in a car: drafting an explanation consumes time and compute, yet trajectory planning must execute several times per second. Nvidia’s answer is adaptive reasoning. On straightforward roads, the model emits a special token in place of reasoning and jumps directly to trajectory generation. When faced with a pedestrian, an oncoming bus, or a traffic controller, it generates “Chain-of-Causation” reasoning, linking its driving decision back to its scene-level causes.

The model was distilled from its larger counterpart, Alpamayo 2 Super, which generated the reasoning traces used as training examples. Nvidia then trained it through closed-loop reinforcement learning inside its AlpaGym simulator, where the model contends with the consequences of its own driving choices. According to the technical blog post, this step delivers a 48.9% boost in driving performance over the strongest checkpoint that skipped it.

160 to 190 ms in the Car, 16 GB on a PC

Downloaded out of the box in BF16, the model runs via PyTorch on an Nvidia GPU: the model card records a peak memory usage of 12 GiB on an H100 and recommends at least 16 GB. For embedded real-time execution, Nvidia outlines a deployment recipe (FP8 calibration, TensorRT engines, DFlash2 speculative decoding) achieving a latency of 160 to 190 ms from image capture to trajectory output on DRIVE AGX Thor, depending on reasoning length. The model can also answer free-form questions about driving scenes.

Alongside this, Nvidia released AlpaJudge, a model that scores the quality of driving reasoning and is calibrated on human judgments. It is used to benchmark models, curate training data, and provide reward signals during post-training.

What Remains Unverified

All reported performance figures originate from Nvidia. The model card states an average of 15.15 km traveled between at-fault incidents (road departures or collisions) across 875 simulated scenarios, and the blog post claims this is roughly 40 times better than Qwen-Drive-1.0, a 4-billion-parameter model, within Nvidia’s AlpaSim simulator. These evaluations were conducted entirely in simulation using Nvidia’s datasets; they offer no guarantees regarding open-road safety. Endorsements from partners (PORT, Niulinx, Embedl) rely on internal testing. Finally, the examples in the repository depend on Nvidia’s PhysicalAI-AV dataset, which comes with its own access terms.