license: other
license_name: embedl-software-2.1
license_link: https://embedl.com/legal/software-licence/v2.1
base_model: nvidia/Alpamayo2-Edge
pipeline_tag: robotics
tags:
- autonomous-driving
- vision-language-action
- alpamayo
- jetson-thor
- tensorrt-edgellm
- nvfp4
- mxfp8
- embedl
Alpamayo 2 Edge — 92 ms on Jetson AGX Thor
10.9 decisions/s on NVIDIA Jetson AGX Thor: 6.6× faster than NVIDIA's PyTorch BF16, with lower trajectory error.
▶ Run Alpamayo 2 Edge live on a real Jetson AGX Thor — free, in your browser: embedl.com/nvidia/alpamayo2-edge/thor
NVIDIA's Alpamayo 2 Edge, the vision-language-action model for autonomous driving, quantized and optimized by Embedl to run in real time on NVIDIA Jetson AGX Thor with TensorRT Edge-LLM, using 4.2× less energy per decision than NVIDIA's PyTorch BF16 release on the same device.
Run it live
Try the optimized model on a real Jetson AGX Thor in Embedl's device farm, for free, with no hardware or setup: Run Alpamayo 2 Edge on AGX Thor →
Get the model
The model and its runtime are distributed as container images from Embedl: see Alpamayo 2 Edge on embedl.com for the downloads and instructions.
Quantization
| Component | Precision |
|---|---|
| Language model | NVFP4 (driving-calibrated), FP16 lm_head |
| Vision encoder (ViT) | MXFP8 |
| Action expert | MXFP8 |
Built from NVIDIA's released weights, nvidia/Alpamayo2-Edge at b92b6aa3.
Runtime: TensorRT Edge-LLM 0.10.1 with Embedl optimizations.
Results on Jetson AGX Thor
| Metric | Embedl NVFP4 / MXFP8 | TensorRT Edge-LLM FP16 | NVIDIA PyTorch BF16 |
|---|---|---|---|
| Latency per decision, median | 91.9 ms | 489.3 ms | 603.7 ms |
| Latency per decision, p90 | 140.3 ms | 840.6 ms | |
| Decisions per second | 10.9 Hz | 1.7 Hz | |
| Energy per decision | 12.45 J | 62.02 J | 51.87 J |
| Power while deciding, mean | 117.6 W | 74.5 W | |
| Power while deciding, peak | 150.2 W | 137.9 W | |
| Model size | 2.37 GiB | 4.95 GiB | |
| Runtime container size | 1.07 GiB | 11.35 GiB | |
| minADE₆ at 5 s (lower is better) | 0.464 m | 0.508 m |
Board power at idle: 30.0 W.
Methodology
- Latency: one PhysicalAI-AV clip, all 180 decisions (one per 0.1 s of driving), run one after another on the device. A decision is the model's whole call from camera frames to trajectory, with the inputs on the GPU. Decisions per second is 1000 / the median.
- Accuracy: minADE₆ at 5 s, the closest of 6 sampled trajectories, average displacement over the first 5 s, on 378 held-out PhysicalAI-AV validation clips, 2 seeds. The Embedl build is measured on the device; NVIDIA's BF16 on an H200 with NVIDIA's own code, on the same events and seeds.
- Power: the board's input power (tegrastats, every 500 ms) sampled during the 180 timed decisions; energy per decision is the mean power times the mean decision time.
- Size: the runtime container image and the model files on disk.
- TensorRT Edge-LLM FP16: the same release in FP16 on the same runtime and device, without Embedl's quantization, caches or speculation.
- NVIDIA PyTorch BF16: NVIDIA's released code (NVlabs/alpamayo2-edge) and
BF16 weights with its
--acceltuned path, on the same device and inputs.
License
This optimized model is licensed under the Embedl Software Licence 2.1. It is derived from nvidia/Alpamayo2-Edge; NVIDIA's original weights remain under their own licence.