NVIDIA’s Compact Driving AI Brings Reasoning Into the Car

NVIDIA has released Alpamayo 2 Edge, a compact reasoning vision-language-action model built to plan autonomous driving actions inside a vehicle. The 2B-parameter model targets NVIDIA DRIVE AGX Thor, bringing trajectory planning and scene explanations to onboard compute instead of a datacenter GPU.
That is the key shift: Alpamayo 2 Edge turns surround-camera video into a 6.4-second driving plan while working within vehicle latency and compute limits. It is the in-car member of the Alpamayo family, joining the 34B cloud model Alpamayo 2 Super as an open NVIDIA VLA model designed for a very different operating environment.
A Smaller Model With a Full Driving Pipeline
Alpamayo 2 Edge combines language understanding, visual perception, reasoning, and action prediction in one system. Its 2.45B backbone pairs a 1.96B language model with a 0.41B SigLIP2 encoder, while a 0.2B diffusion action expert sits on top. The backbone comes from NVIDIA Cosmos 3 Edge and is Nemotron-based.
NVIDIA markets the system as 2B, while Hugging Face metadata shows 3B. The weights ship in BF16 and run as a PyTorch model, with workstation use calling for at least 16 GB of GPU memory. NVIDIA measured 12.06 GiB of peak process memory on 1 H100 with 6 trajectory candidates, although other GPU architectures have not yet been validated.
For each observation, the model reads 6 cameras and 22 frames. The 2 front cameras provide 5 frames each, while the other 4 cameras provide 3 frames each. Alpamayo 2 Edge then combines that visual input with a navigation instruction and 20 historical ego poses.
The output is a sequence of 64 waypoints covering the period from 0.1 to 6.4 seconds. Its action expert predicts acceleration and curvature, and a unicycle model integrates those predictions into the vehicle’s trajectory.
Reasoning Only When the Road Demands It
Alpamayo 2 Edge does not produce a full explanation for every routine driving moment. Its text output uses a Chain-of-Causation reasoning trace, called CoC, when a scene needs deeper consideration. During routine driving, the model emits special <NOMINAL> tokens instead.
That selective approach cuts the number of tokens generated for ordinary scenes, which means less compute per planning step. The model can still answer free-form questions about a driving scene, giving it a role beyond trajectory prediction and allowing it to explain what it sees when an unusual situation requires attention.
NVIDIA lists several ingredients behind the model’s training process:
- Offline distillation from CoC autolabels generated by Alpamayo 2 Super
- About 115,000 hours of multi-camera driving video
- About 3.7M CoC traces
- Closed-loop reinforcement learning post-training in AlpaGym, where the model learns from its own actions in simulation
- Multi-resolution inputs that reduce the cost of encoding surround views
This training design connects a large cloud model with a smaller onboard model, then adds simulation-based learning to help Alpamayo 2 Edge make decisions inside a closed loop.
Latency, Simulation Results, and Thor Performance
NVIDIA reports 160 ms to 190 ms of pixel-to-trajectory latency, depending on how much the model reasons. Real-time use on Thor needs the Alpamayo Developer Hub recipe, and NVIDIA’s reference setup uses DRIVE AGX Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4, and TensorRT 11.3.1.9.
In NVIDIA’s demonstration video, the PyTorch BF16 version plans at 0.76 Hz. The optimized FP8 Thor pipeline reaches 13.02 Hz in AlpaSim, while partner Embedl says it reached 10 Hz on AGX Thor in under 48 hours.
The closed-loop results put the edge model against Alibaba’s 4B Qwen-Drive-1.0. In AlpaSim, Alpamayo 2 Edge covers 15.15 km between at-fault incidents, compared with 0.25 km for Qwen-Drive-1.0—about 60 times the distance. Its open-loop minADE measures 0.53 m versus 1.68 m, based on NVIDIA’s own runs.
The model card records 15.15 ± 0.04 km per at-fault incident over 875 NuRec scenarios and 3 rollouts. It also lists an open-loop minADE_6 at 5 seconds of 0.52 m on 1,392 PhysicalAI-AV samples, along with LingoQA at 75.2, SURDS at 86.50, KITScenes scene at 89.46, and nuReasoning numeric at 48.67.
LingoQA gives Alpamayo 2 Edge a 75.2 score, trailing Alpamayo 2 Super at 79.2. That gap highlights the tradeoff NVIDIA is making: the edge model gives up some capability to fit inside the vehicle and deliver plans under onboard latency limits.
The Bigger Autonomous Driving Race
Alpamayo 2 Edge arrives as autonomous driving companies push different pieces of the technology into real-world services. In Las Vegas, Uber has added Zoox robotaxis to its ride-hailing app, and customers can increase their chances of getting a Zoox robotaxi while still declining a robot ride.
Uber CEO Dara Khosrowshahi said the company plans to introduce a dedicated autonomous option in its app as “availability and reliability improve.” Las Vegas will become the first U.S. city where riders can access two fully driverless partners through the Uber app, with Hyundai’s self-driving subsidiary Motional also serving as an Uber partner.
The wider robotaxi market is shifting as well. Tesla has surpassed Waymo in Texas in the number of robotaxis registered with the state, with Tesla at 1,507 and Waymo at 1,345. Waymo has also pulled its robotaxis off freeways again, showing how deployment still depends on careful control of operating conditions.
That makes NVIDIA’s edge strategy especially important. A model that can interpret six camera feeds, reason when needed, and produce a 6.4-second plan at vehicle-ready latency gives autonomous systems a path toward more capable onboard decision-making. Alpamayo 2 Edge is not just a smaller model; it is NVIDIA’s push to make reasoning fit where the driving happens.
Based on
- NVIDIA Releases Alpamayo 2 Edge: A 2B Reasoning VLA for DRIVE Thor — marktechpost.com
- TechCrunch Mobility: A/B testing robots vs. humans | TechCrunch — techcrunch.com
- Mercedes’ clever new driving assist handles complex traffic like a champ | Ars OpenForum — arstechnica.com
- Heterogeneous multimodal driver assistance using a Jetson Nano–ESP32 architecture | Scientific Reports — nature.com
- Waymo suspends freeway driving after California construction zone complaint | Page 3 | Ars OpenForum — arstechnica.com




