Physical AI
Run your own ONNX model on NVIDIA Jetson Thor with accelerated engines built, quantized, and benchmarked by ZETIC Melange.
The Physical AI lane takes an ONNX model you already have and turns it into accelerated engines for NVIDIA Jetson Thor. You upload the model together with a handful of real input samples. Melange freezes the graph to your input shapes, builds one engine per precision, measures every engine against your original FP32 model, and publishes the result as a benchmark scorecard.
Those samples do three jobs at once, which is why they matter more here than anywhere else in Melange: they pin the static shapes, they calibrate quantization, and they are the data every accuracy number is measured on. Getting them right is most of the work, so they have a page of their own.
Early Access Program
Physical AI is available only to members of the ZETIC Early Access Program. If Physical AI Model does not appear in the upload dialog, your account is not enrolled — contact us to request access.
How it differs from the mobile lane
| Mobile (Android / iOS) | Physical AI | |
|---|---|---|
| Target hardware | Phone NPU, GPU, CPU | NVIDIA Jetson Thor (Linux aarch64) |
| Model source | .pt2 or .onnx, or a Hugging Face repository | .onnx you upload |
| Input samples | .npy, one file per input | .npz, one archive per sample |
| Runtime SDK | Android, iOS, Flutter | Python, C++ (Early Access) |
Projects, model keys, and personal keys work exactly as they do everywhere else in Melange.
What Melange builds
Every upload is converted at four precisions:
| Precision | What it is |
|---|---|
| FP16 | Half precision. No calibration, no accuracy cliff — treat it as the baseline. |
| FP8 | 8-bit floating point, calibrated on your samples. |
| INT8 | 8-bit integer, calibrated on your samples. |
| INT4 | 4-bit weight-only quantization. |
Lower precision buys size, and sometimes speed. It costs accuracy. How much of each is exactly what the scorecard answers — for your model, on your data, rather than in general. A variant that fails to build or benchmark appears in the table without numbers rather than disappearing quietly.
The workflow
Export to ONNX
One .onnx file per module. If your model is a pipeline, upload each stage as its own model.
Build a sample set
One .npz per sample, with array keys matching your model's input names. Use real data — this is what calibration and scoring read.
Upload
Choose Physical AI Model in the upload dialog, attach the .onnx and the samples, and wait for the run to finish.
Read the scorecard
The model report gains a Benchmark table: latency, host memory, engine size, and output SNR for every precision that built.
Run on Jetson Thor
The Python and C++ runtimes ship through the Early Access Program rather than a public package index — contact us to get them along with your converted engines.
Requirements
- A Melange account enrolled in the Early Access Program, and a personal key (how to get one)
- A single
.onnxfile, plus its external data files if the model stores weights outside the graph - At least one
.npzsample; up to 50 - An NVIDIA Jetson Thor device to run the result on