BYOM

Bring your own model — upload a pre-trained artifact, validate it, and register it as a model you can deploy.

Already trained a model elsewhere? BYOM (bring your own model) puts it in your catalog in three calls: upload the file, validate it (the platform detects the framework, infers the input/output signature, and runs a live forward pass), then register it as a kind: "byom" model.

Accepted formats: PyTorch (.pt), TensorFlow/Keras (.h5, SavedModel), ONNX (.onnx), and scikit-learn / XGBoost pickles (.pkl).

Upload size is capped by your plan (the byom_max_mb limit — for example Starter allows files up to 500 MB, Team up to 2 GB, and Enterprise has no cap; the free tier doesn't include BYOM). The cap is resolved from your plan automatically on every call. Uploads also charge credits per MB — see Billing for tiers and rates.

Step 1 — Upload

Send the artifact as a multipart file field. The upload is charged per MB before storage (and refunded automatically if storage fails):

curl -X POST "$API_BASE/v1/models/byom/upload" \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Workspace-Id: $WORKSPACE_ID" \
  -F "file=@./sales_forecaster.onnx"

from predictai import PredictAI

client = PredictAI(token="pa_live_…", workspace_id="ws_…")

data = client.byom.post_models_byom_upload(files={"file": open("./sales_forecaster.onnx", "rb")})

import { PredictAI } from "@predictai/sdk";
import { openAsBlob } from "node:fs";

const client = new PredictAI({ token: "pa_live_…", workspaceId: "ws_…" });

const form = new FormData();
form.append("file", await openAsBlob("./sales_forecaster.onnx"), "./sales_forecaster.onnx");

const data = await client.byom.postModelsByomUpload({ form });

{
  "data": {
    "model_id": "1f4c3a9e6b8d4e219c7a5b3f8d2e6a41",
    "model_uri": "s3://celmind-byom/ORG_ID/1f4c3a9e…/v1/model.onnx",
    "local_staging_path": "/tmp/byom_1f4c3a9e….onnx",
    "size_bytes": 48211284,
    "ext": ".onnx",
    …
  }
}

Keep the model_id — the next two calls take it.

Request

Field	Meaning
`file` required	The model artifact, as a multipart form field. One file per upload.

Errors

Status	Why	Example message
`400`	No file in the request	`"multipart 'file' field required"`
`400`	Empty file	`"file is empty"`
`402`	Not enough credits for the per-MB upload charge	`error.code` `insufficient_credits`, `error.details.needed` `46.0`
`403`	Your plan doesn't include BYOM	`error.code` `plan_limit`
`413`	File exceeds your plan's size cap	`error.code` `plan_limit`, `error.details.cap_mb` `500`

Step 2 — Validate

Validation auto-detects the framework, infers the model's signature, and runs one forward pass with random input of the inferred shape. It's a paid operation (byom.validate):

curl -X POST "$API_BASE/v1/models/byom/validate" \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Workspace-Id: $WORKSPACE_ID" \
  -H "Content-Type: application/json" \
  -d '{ "model_id": "1f4c3a9e6b8d4e219c7a5b3f8d2e6a41" }'

from predictai import PredictAI

client = PredictAI(token="pa_live_…", workspace_id="ws_…")

data = client.byom.post_models_byom_validate(json={"model_id": "1f4c3a9e6b8d4e219c7a5b3f8d2e6a41"})

import { PredictAI } from "@predictai/sdk";

const client = new PredictAI({ token: "pa_live_…", workspaceId: "ws_…" });

const data = await client.byom.postModelsByomValidate({ json: { model_id: "1f4c3a9e6b8d4e219c7a5b3f8d2e6a41" } });

{
  "data": {
    "model_id": "1f4c3a9e6b8d4e219c7a5b3f8d2e6a41",
    "detection": { "framework": "onnx", "confidence": 0.98, "notes": "ONNX graph header" },
    "signature": {
      "input_signature":  [ { "name": "input",  "shape": [-1, 20, 3], "dtype": "float32" } ],
      "output_signature": [ { "name": "output", "shape": [-1, 14, 1], "dtype": "float32" } ]
    },
    "validation": {
      "ok": true,
      "framework": "onnx",
      "file_size_bytes": 48211284,
      "static_warnings": [],
      "static_errors": [],
      "dynamic_passed": true,
      "forward_latency_ms": 11.4,
      "output_shape": [1, 14, 1],
      "notes": ""
    }
  }
}

Request body

Field	Meaning
`model_id` required	The ID from upload.
`framework` optional	Skip auto-detection: `onnx`, `pytorch`, `tensorflow`, `sklearn`, `xgboost`.
`input_shape_hint` optional	An input shape like `[1, 20, 3]` — helps when signature inference is ambiguous.

Response fields

Field	Meaning
`detection`	The detected `framework` with a `confidence` score and what gave it away.
`signature`	The inferred `input_signature` / `output_signature` — name, shape, dtype per tensor. You pass these back at register.
`validation`	The verdict: `ok`, `static_warnings` / `static_errors`, whether the live forward pass ran (`dynamic_passed`), and its latency.

A model that fails validation still returns 200 — check validation.ok and read static_errors:

"static_errors": ["File size 812.4 MB exceeds your plan's cap of 500 MB"]

Pickle files get a standing warning: they execute code at load time, so consider re-exporting to ONNX.

Errors

Status	Why	Example message
`400`	No `model_id` in the body	`"model_id required"`
`404`	Upload expired or wrong id	`"No staged file found for model_id 1f4c…"`
`402`	Not enough credits	`error.code` `insufficient_credits`, `error.details.operation` `"byom.validate"`

Step 3 — Register

Pass the signature and validation objects from step 2 back in, plus a name. This creates the catalog entry:

curl -X POST "$API_BASE/v1/models/byom/register" \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Workspace-Id: $WORKSPACE_ID" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "1f4c3a9e6b8d4e219c7a5b3f8d2e6a41",
    "model_uri": "s3://celmind-byom/ORG_ID/1f4c3a9e…/v1/model.onnx",
    "framework": "onnx",
    "name": "Sales forecaster (imported)",
    "description": "Trained offline on 3 years of daily sales",
    "visibility": "private",
    "signature": { "input_signature": […], "output_signature": […] },
    "validation": { "ok": true, … },
    "segment_id": "SEGMENT_ID"
  }'

from predictai import PredictAI

client = PredictAI(token="pa_live_…", workspace_id="ws_…")

data = client.byom.post_models_byom_register(json={
    "model_id": "1f4c3a9e6b8d4e219c7a5b3f8d2e6a41",
    "model_uri": "s3://celmind-byom/ORG_ID/1f4c3a9e…/v1/model.onnx",
    "framework": "onnx",
    "name": "Sales forecaster (imported)",
    "description": "Trained offline on 3 years of daily sales",
    "visibility": "private",
    "signature": {"input_signature": […], "output_signature": […]},
    "validation": {"ok": True, …},
    "segment_id": "SEGMENT_ID",
})

import { PredictAI } from "@predictai/sdk";

const client = new PredictAI({ token: "pa_live_…", workspaceId: "ws_…" });

const data = await client.byom.postModelsByomRegister({
  json: {
    model_id: "1f4c3a9e6b8d4e219c7a5b3f8d2e6a41",
    model_uri: "s3://celmind-byom/ORG_ID/1f4c3a9e…/v1/model.onnx",
    framework: "onnx",
    name: "Sales forecaster (imported)",
    description: "Trained offline on 3 years of daily sales",
    visibility: "private",
    signature: { input_signature: […], output_signature: […] },
    validation: { ok: true, … },
    segment_id: "SEGMENT_ID",
  },
});

{
  "data": {
    "blueprint_id": "b7d2e911-…",
    "model_id": "5a80c4f3-…",
    "training_id": "e2c619ab-…",
    …
  }
}

Request body

Field	Meaning
`model_id` required	The ID from upload.
`model_uri` required	The storage URI from the upload response, echoed back.
`framework` required	The framework from the validation response.
`name` required	Display name for the catalog entry.
`signature` required	The `signature` object from validate, passed back unchanged.
`validation` required	The `validation` object from validate, passed back unchanged. `ok` must be `true`.
`segment_id` optional	Bind the model to a segment: also creates a pipeline and a ready-to-promote training record (see below).
`description` optional	Free-text description.
`visibility` optional	`private` (default) or `public`.
`license` optional	Defaults to `proprietary`.
`preprocessing` optional	Preprocessing applied before inference.

Response fields

Field	Meaning
`blueprint_id`	The new model's catalog ID (response fields use the internal name, like the route paths).
`model_id`	Only when `segment_id` was passed: the ID of the pipeline created to bind the model to that segment.
`training_id`	Only when `segment_id` was passed: a ready-to-promote training record — go straight to deploy with one `POST /v1/deployments` call.

Omit segment_id and you get just the catalog entry.

Compatibility with your segment

When you pass a segment_id, the platform checks the model's signature against the segment's layout before saving anything: the number of input features must match the segment's features (base + engineered), and the output must map onto the segment's labels. Mismatches return a 422 with the concrete issue — never a broken deployment later:

{
  "error": {
    "code": "byom_incompatible_segment",
    "message": "The model's signature doesn't match the segment's feature/label layout.",
    "details": {
      "issues": ["Model expects 5 input features per step but the segment provides 3 (revenue, units, price)."],
      "expected": { "n_features": 3, "features": ["revenue", "units", "price"], "n_labels": 1, "labels": ["revenue"] }
    }
  }
}

The same check runs when you later create a pipeline from a registered BYOM model (POST /v1/pipelines) and again as a backstop at promotion time — so a segment whose features changed since registration is caught before it can break serving.

Errors

Status	Why	Example message
`400`	Required field missing	`"Missing required fields: signature, validation"`
`403`	Plan quota for uploaded models reached	`"BYOM model limit reached. Your plan allows 3 uploaded models per workspace."`
`404`	Segment not found in this workspace	`"Segment SEGMENT_ID not found in this workspace"`
`422`	Signature doesn't fit the segment	`error.code` `byom_incompatible_segment`
`422`	Artifact can't be converted for serving	`error.code` `byom_staging_failed`

BYOM models don't count against the regular per-workspace model quota — they have their own byom_models quota, checked here at register time. See Billing.

Download your model

Your model is never locked in. Fetch a time-limited download link for the exact file you uploaded:

curl "$API_BASE/v1/models/byom/BLUEPRINT_ID/download" \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Workspace-Id: $WORKSPACE_ID"

from predictai import PredictAI

client = PredictAI(token="pa_live_…", workspace_id="ws_…")

data = client.byom.get_models_byom_by_blueprint_id_download("BLUEPRINT_ID")

import { PredictAI } from "@predictai/sdk";

const client = new PredictAI({ token: "pa_live_…", workspaceId: "ws_…" });

const data = await client.byom.getModelsByomByBlueprintIdDownload("BLUEPRINT_ID");

{
  "data": {
    "download_url": "https://celmind-byom.s3.amazonaws.com/…&X-Amz-Expires=3600…",
    "filename": "9f2c41d8-7b3a-4e1f-8c2d-5a6b7c8d9e0f.onnx",
    "framework": "onnx",
    "expires_in_seconds": 3600
  }
}

The URL is valid for one hour; request a fresh one any time. The file saves under the model's ID (the download link carries a Content-Disposition header, so the name is applied automatically).

Errors

Status	Why	Example message
`403`	Not your model (and not public)	`"You don't have access to this model"`
`404`	Unknown id, or not a BYOM model	`"BYOM model BLUEPRINT_ID not found"`

On this page