Is Alpamayo 2 Edge state of the art?Pending benchmarks.Announced Nov 10, 2026; no independent results yet.
NVIDIA's 2B-parameter reasoning vision-language-action model in the Alpamayo family, built on the Nemotron-based backbone of NVIDIA Cosmos 3 Edge and paired with a diffusion action expert. It generates Chain-of-Causation reasoning only when needed and produces driving trajectories from surround-camera observations, temporal context, and navigation inputs. Released November 10, 2026 as part of the Alpamayo Open Platform, with public weights on Hugging Face under OpenMDW-1.1 and code under Apache License 2.0. Sized for real-time in-vehicle use on NVIDIA DRIVE AGX Thor.
The details so far
- Lab
- NVIDIA
- Announced
- Nov 10, 2026
- API
- Not yet · No public cloud API announced
- Open weights
- Published · Public, ungated Hugging Face repo nvidia/Alpamayo2-Edge (BF16 safetensors). Official card release date November 10, 2026
- Parameters
- NVIDIA describes it as a 2B-parameter model. Hugging Face card: backbone 2.45B (1.96B language model, 0.41B SigLIP2 vision encoder, 0.08B projector) plus a 0.2B action expert
- Type
- Reasoning vision-language-action (VLA) model with adaptive Chain-of-Causation reasoning and a diffusion action expert (NVIDIA blog and model card)
- Base
- Built on the 2B Nemotron-based reasoning backbone of NVIDIA Cosmos 3 Edge (NVIDIA blog and model card)
- Target hardware
- NVIDIA DRIVE AGX Thor (NVIDIA blog)
- Inputs
- Multi-camera video, ego-motion history, and navigation instructions; optional free-form visual question answering (NVIDIA blog and model card)
- Trajectory
- 64 waypoints from 0.1 through 6.4 seconds at 0.1-second intervals (model card)
- Training data
- Roughly 115,000 hours of multi-camera driving video and about 3,700,000 Chain-of-Causation traces (model card)
- License
- Model weights OpenMDW-1.1; source code Apache License 2.0 (model card and NVlabs/alpamayo2-edge)
- Code
- https://github.com/NVlabs/alpamayo2-edge
- Related
- Separate from Alpamayo 2 Super (used as a teacher for CoC autolabels), AlpaJudge, Cosmos 3 Edge, and other Cosmos pages
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: post-trained with closed-loop reinforcement learning in NVIDIA AlpaGym, improving closed-loop driving performance by 48.9% on top of the best-performing checkpoint without closed-loop RL (NVIDIA blog).
Lab claim: optimized pixel-to-trajectory inference latency of 160 ms to 190 ms on NVIDIA DRIVE AGX Thor, depending on the amount of reasoning (NVIDIA blog).
Lab claim: Lingo-Judge score of 75.2 on LingoQA (NVIDIA Hugging Face model card).
Lab claim: in AlpaSim closed-loop evaluation on 875 scenarios from PhysicalAI-AV-NuRec, with three rollouts per scenario, 15.15 ± 0.04 km driven per at-fault incident (NVIDIA Hugging Face model card).
Lab claim: open-loop minADE_6 at 5s of 0.52 m on 1,392 samples from the PhysicalAI-AV dataset (NVIDIA Hugging Face model card).
Lab claim: achieves approximately 40× the distance between at-fault incidents in AlpaSim compared with Qwen-Drive-1.0 (NVIDIA blog).
Sources
LABNVIDIA blog: Running AV Reasoning Models Within Onboard Compute and Latency Limitsdeveloper.nvidia.com/blog/running-av-reasoning-models-within-onboard-compute-and-latency-limits/ · Nov 10, 2026REPOnvidia/Alpamayo2-Edge on Hugging Facehuggingface.co/nvidia/Alpamayo2-Edge · Nov 10, 2026REPONVlabs/alpamayo2-edge on GitHubgithub.com/NVlabs/alpamayo2-edgeTracking continues until an independent board lists it.