Output & manifest - R3AL.AI

The deliverable

When quantize() or train_qat() finishes on the platform, the SDK downloads and unpacks the bundle into your output_dir:

output_dir/
├── model.quantized.onnx      ← quantized model (stem matches source;
│                               QAT jobs write model_qat.quantized.onnx)
└── r3alai_manifest.json      ← proof + full config, shipped inside the bundle

result.path points at this directory, so local load() and benchmark() work on it directly.

The manifest

r3alai_manifest.json ships inside every downloaded bundle. It is the customer-facing proof that a model was processed by R3AL.AI:

{
  "quantized": true,
  "producer": "R3AL.ai Quant SDK",
  "format": "onnx",
  "source_model": "model.onnx",
  "output_model": "model.quantized.onnx",
  "method": "ptq_dynamic",
  "config": {
    "method": "ptq_dynamic",
    "bits": 8,
    "mode": "ptq"
  },
  "domain": "production-v1",
  "created_at": "2026-07-19T12:00:00+00:00"
}

Field	Meaning
`quantized` / `producer`	Confirms the file came from the R3AL engine
`method`	The method used
`config`	The full config used for the job
`domain`	Optional free-form tag you pass with the job
`created_at`	UTC timestamp

Embedded ONNX metadata

The quantized ONNX file itself is stamped with metadata properties, so provenance survives even if the manifest is separated from the model:

import onnx

model = onnx.load("model.quantized.onnx")
print({p.key: p.value for p in model.metadata_props})
# {'r3al_quantized': 'true', 'r3al_method': 'ptq_dynamic', 'r3al_bits': '8'}

Loading a quantized bundle

The SDK resolves the manifest for you when loading (locally, via onnxruntime):

from r3alai.quant import Quantizer

bundle = Quantizer().load("./out")
session = bundle["session"]      # onnxruntime.InferenceSession
manifest = bundle["manifest"]    # parsed manifest dict
bundle["path"]                   # path to the resolved .onnx file