Jobs: upload, quantize, qat, download

Quantization jobs run on R3AL GPU infrastructure. The flow is always: upload the model, submit the job (which returns a job_id), poll the job, then download its artifacts.

Upload a model

Models are uploaded to R2 with a presigned PUT. Ask for a grant, then PUT the file bytes to the returned URL.

curl -X POST https://platform.r3al.ai/v1/uploads \
  -H "Authorization: Bearer r3l_live_..." \
  -H "Content-Type: application/json" \
  -d '{"filename": "model.onnx", "content_type": "application/octet-stream", "size_bytes": 10485760}'

{ "upload_url": "https://...presigned...", "object_key": "uploads/ab12/model.onnx" }

PUT the file to upload_url, then reference it in a job as r2://<object_key>. Single uploads are limited to 5 GB. Calibration images are uploaded the same way.

Submit a job

Jobs accept r2://<key> model references and always return 202 with a job_id.

Endpoint	Purpose	Notes
`POST /v1/quantize`	PTQ: model in, quantized ONNX out	`method`: `ptq_static` (default), `ptq_dynamic`
`POST /v1/qat`	QAT: model + `calibration_data`, train, quantized ONNX out	Always async

curl -X POST https://platform.r3al.ai/v1/quantize \
  -H "Authorization: Bearer r3l_live_..." \
  -H "Content-Type: application/json" \
  -d '{"model": "r2://uploads/ab12/model.onnx", "method": "ptq_static", "calibration_data": ["r2://uploads/ab12/img1.jpg"]}'

{ "job_id": "job_abc123", "status": "pending" }

Both quantize and qat are asynchronous: they return a job_id immediately, since a run can take longer than an HTTP client’s timeout. Poll the job for progress and the result.

Poll a job

curl https://platform.r3al.ai/v1/jobs/job_abc123 \
  -H "Authorization: Bearer r3l_live_..."

{
  "job_id": "job_abc123",
  "status": "running",
  "progress": 62.0,
  "progress_message": "calibrating",
  "result": null
}

status moves through pending, running, then a terminal completed or error. On completed, result carries the deliverable metadata; on error, an error object with a code and message.

Download artifacts

Once a job is completed, request presigned download URLs for its artifacts.

curl https://platform.r3al.ai/v1/jobs/job_abc123/artifacts \
  -H "Authorization: Bearer r3l_live_..."

{
  "artifacts": [
    { "filename": "model.quantized.zip", "url": "https://...presigned-get..." }
  ]
}

Download each url (presigned GET). The bundle contains the quantized ONNX model and r3alai_manifest.json.

Delete a job

curl -X DELETE https://platform.r3al.ai/v1/jobs/job_abc123 \
  -H "Authorization: Bearer r3l_live_..."

Removes the job and its artifacts. 404 if the id does not exist.

Usage and plan

curl https://platform.r3al.ai/v1/usage \
  -H "Authorization: Bearer r3l_live_..."

Returns your plan and remaining free runs. Each quantize or qat job consumes one run; a run beyond your quota fails with 402 plan_limit_reached (see Errors).

Method field

method applies to POST /v1/quantize only:

ptq_static (default): requires calibration_data (a list of uploaded image refs).
ptq_dynamic: no calibration data needed; it quantizes MatMul and Gemm layers, so it suits MatMul-heavy models rather than convolution-heavy backbones.

QAT settings (wbit, abit, epochs) go on POST /v1/qat. See the Python SDK for the field-level reference; only the transport differs.

Get Started

Endpoints

Reference

Jobs: upload, quantize, qat, download

Upload a model

Submit a job

Poll a job

Download artifacts

Delete a job

Usage and plan

Method field

​Upload a model

​Submit a job

​Poll a job

​Download artifacts

​Delete a job

​Usage and plan

​Method field

Upload a model

Submit a job

Poll a job

Download artifacts

Delete a job

Usage and plan

Method field