Product Launches
NVIDIA Releases Alpamayo 2 Edge, a 2B-Parameter Driving Model That Plans in 160-190 ms
By News Desk ·
Key takeaways
- Alpamayo 2 Edge is a 2B-parameter vision-language-action model that turns six-camera video into a 6.4-second, 64-waypoint plan.
- NVIDIA reports 160 ms to 190 ms pixel-to-trajectory latency on DRIVE AGX Thor with the FP8 TensorRT Edge-LLM recipe.
- In AlpaSim, the model covered 15.15 km per at-fault incident against 0.25 km for Alibaba's 4B Qwen-Drive-1.0.
NVIDIA released Alpamayo 2 Edge, a 2B-parameter vision-language-action model that plans driving trajectories and explains them inside the vehicle, targeting the DRIVE AGX Thor platform rather than a datacenter GPU. MarkTechPost reported the release on November 22, 2026, with weights under the OpenMDW-1.1 license and code under Apache 2.0.
What Alpamayo 2 Edge Does In The Car
NVIDIA built the model to turn surround-camera video into a 6.4-second plan. For trajectory prediction it reads 6 cameras and 22 frames per observation, plus a navigation instruction and 20 historical ego poses, and outputs 64 waypoints from 0.1 to 6.4 seconds. A 0.2B diffusion action expert predicts acceleration and curvature, which a unicycle model integrates.
The model card describes a 2.45B backbone pairing a 1.96B language model with a 0.41B SigLIP2 encoder, drawn from NVIDIA Cosmos 3 Edge and Nemotron-based, while Hugging Face metadata shows 3B. During routine driving the model emits special <NOMINAL> tokens instead of a full reasoning trace, writing a Chain-of-Causation trace only when a scene needs it. NVIDIA reports 160 ms to 190 ms pixel-to-trajectory latency on DRIVE AGX Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4 and TensorRT 11.3.1.9.
In AlpaSim closed-loop driving, the model covered 15.15 km per at-fault incident over 875 NuRec scenarios with 3 rollouts, against 0.25 km for Alibaba's 4B Qwen-Drive-1.0, both in NVIDIA's own runs. Its LingoQA score of 75.2 trails Alpamayo 2 Super at 79.2. NVIDIA also released AlpaJudge, a human-calibrated judge for reasoning traces that lifted reasoning quality 10.2% as an RL reward by judge and 11.4% by human evaluation.
NVIDIA's Funding And Deal Record
According to akta.pro company records, NVIDIA issued $25 billion in high-grade bonds in June 2026 to expand liquidity and fund a 2,400% dividend increase. The same records show a January 2026 acquisition of inference chip startup Groq for $20 billion and a December 2025 acquisition of SchedMD, whose Slurm workload manager is used on about 60% of the world's supercomputers.
Earlier Moves In NVIDIA's Physical AI Push
In June 2026, a roughly 30-member LG Group delegation held high-level talks at NVIDIA's Santa Clara headquarters on AI infrastructure, robotics and Physical AI cooperation, according to akta.pro company records. The same month, Argentum AI signed approximately $7.8 billion in agreements to deploy around 47,000 NVIDIA GB300 GPUs across a 300-megawatt datacenter buildout.
Where To Find The Model
Weights ship in BF16 and run as a PyTorch model, with real-time use on Thor requiring the Alpamayo Developer Hub recipe that calibrates FP8 scales and builds TensorRT engines. Inference code is on GitHub, and the model supports fine-tuning.