Tasks and inputs
Alpamayo 2 Edge is listed for Image-text-to-text, Multimodal. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Alpamayo 2 Edge is NVIDIA's open 2B-class driving VLA that outputs trajectories with causal explanations, sized for DRIVE AGX Thor in-vehicle use.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; no per-token price is published. Your cost is the GPU time you run it on.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Explainable planning has mostly been a datacentre luxury in autonomous driving. Alpamayo 2 Edge tries to put it in the car. It is a compact reasoning vision-language-action model that predicts where the vehicle should go and writes a Chain-of-Causation explanation of why, and it is meant to be fine-tuned and then deployed on NVIDIA DRIVE AGX Thor.
Trajectory prediction looks at six cameras across 22 frames: five each from the front wide and front tele cameras over the last two seconds, and three each from the four side and rear cameras. Add a text navigation instruction and 20 past ego poses at 0.1-second steps.
Out comes a trajectory of 64 waypoints covering 0.1 to 6.4 seconds, each with an XYZ position and a 3x3 rotation, plus reasoning text. Sample several candidates and each gets its own reasoning. The same model also answers free-form questions about a scene, using four frames per camera.
| Part | Size |
|---|---|
| Language model (from Cosmos 3 Edge, Nemotron-based) | 1.96B |
| SigLIP2 vision encoder | 0.41B |
| Projector | 0.08B |
| Diffusion action expert | 0.2B |
Weights are BF16 under OpenMDW-1.1. The code is Apache-2.0.
enable_thinking=False; the driving prompt supplies its own empty reasoning block.The repository dates from 12 August 2026, but NVIDIA lists 10 November 2026 as the public release.
Replay logged drives with the camera layout remapped to the expected six IDs. Score open-loop displacement error at 1, 3 and 6 seconds, then have an engineer read fifty explanations for cases where the plan was wrong. An explanation that sounds right on a bad trajectory is the failure mode to look for. Only then consider closed-loop simulation.
It is part of NVIDIA's open platform for vehicles; related releases are on the NVIDIA provider page, and broader vision-language models are under multimodal.
Compare the model on the work you need to complete.
Alpamayo 2 Edge is listed for Image-text-to-text, Multimodal. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under OpenMDW-1.1 (weights); Apache-2.0 (code). Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
shanghai-ai-lab
Sage-SciPerception is Shanghai AI Lab's set of projectors feeding crystals, spectra, biological sequences and biosignals into a language model's input.
inclusionAI
Ling 3.2 Flash is inclusionAI's 560B-A25B multimodal MoE with a 1M-token window, built for visual agents, 3D tasks and medical image reasoning.
Swiss AI
Apertus 1.5 8B is the Swiss AI Initiative's open multilingual model, now taking text, images and audio with a 262K-token context window.
kyutai
REKI-2B is Kyutai's open 2.4B robot vision-language-action model with cheap register-token memory, pretrained only on public data.
black-forest-labs
FLUX 3 Action is Black Forest Labs' open 7B world action model for robots, predicting action chunks and future video from cameras, state and text.