ML Model Repositories in JFrog Artifactory¶

← Back to JFrog AI & ML

Machine learning models are large binary files — .onnx, .gguf, safetensors, .pkl — that need the same versioning, access control, and traceability you apply to software artifacts. JFrog Artifactory's Generic repositories provide exactly that.

All steps use JFrog SaaS at https://<company>.jfrog.io.

Why Store Models in Artifactory?¶

Problem	Artifactory Solution
Models stored in ad-hoc S3 buckets	Centralized, versioned storage with metadata
No audit trail of which model is in production	Build info links model version to training run
Untrusted models pulled from Hugging Face	Xray scans; Curation blocks unvetted models
Different teams download same model repeatedly	Proxy/cache model registries — download once
Model rollback is manual	Promote/demote between `staging` and `prod` repos

Step 1: Create a Generic Local Repository for Models¶

Go to Administration → Repositories → + New Repository
Select Local
Choose Generic
Set Repository Key: ml-models-local
Click Create Local Repository

Step 2: Create a Remote Repository — Hugging Face Hub Proxy¶

JFrog can proxy Hugging Face model downloads and cache them locally:

Go to Administration → Repositories → + New Repository
Select Remote
Choose Generic
Set Repository Key: huggingface-remote
Set URL: https://huggingface.co
Click Create Remote Repository

Step 3: Upload a Model via JFrog CLI¶

# Upload a trained PyTorch model
jf rt upload my-model.pt ml-models-local/nlp-classifier/1.0.0/

# Upload with rich metadata
jf rt upload my-model.onnx ml-models-local/image-detector/2.3.0/ \
  --props "framework=pytorch;task=image-classification;accuracy=0.94;dataset=imagenet;training-run=train-#42"

# Upload multiple model files
jf rt upload "models/*.gguf" ml-models-local/llm/llama3-8b/v1/

Step 4: Download a Model in Inference Code¶

# Download specific model version
jf rt download "ml-models-local/nlp-classifier/1.0.0/my-model.pt" ./models/

# Download by property (find production model)
jf rt download \
  --props "env=production;task=image-classification" \
  "ml-models-local/*" ./inference/models/

In Python (using the JFrog REST API):

import requests
import os

JFROG_URL = "https://<company>.jfrog.io/artifactory"
TOKEN = os.environ["JFROG_TOKEN"]

response = requests.get(
    f"{JFROG_URL}/ml-models-local/nlp-classifier/1.0.0/my-model.pt",
    headers={"Authorization": f"Bearer {TOKEN}"},
    stream=True,
)

with open("model.pt", "wb") as f:
    for chunk in response.iter_content(chunk_size=8192):
        f.write(chunk)

Step 5: Tag Model Properties for Discovery¶

Properties (key-value metadata) make models searchable and auditable:

# Add properties after upload
jf rt set-props \
  "ml-models-local/nlp-classifier/1.0.0/my-model.pt" \
  "env=staging;approved-by=alice;approval-date=2026-02-21"

# Search for all production-approved models
jf rt search \
  --props "env=production" \
  "ml-models-local/**/*.pt"

Recommended Folder Structure¶

ml-models-local/
├── nlp/
│   ├── sentiment-classifier/
│   │   ├── 1.0.0/
│   │   │   └── sentiment.onnx
│   │   └── 2.0.0/
│   │       └── sentiment.onnx
├── vision/
│   └── image-detector/
│       └── 3.1.0/
│           └── yolo-v8.pt
└── llm/
    └── llama3-8b/
        └── v1/
            └── llama3-8b-q4.gguf

Use Cases¶

Scenario	Solution
Training pipeline outputs model file	Upload to `ml-models-local` with build info
Inference service needs latest approved model	Query by `env=production` property
Roll back to previous model version	Re-promote an older model file to production path
Scan model's Python dependencies for CVEs	Attach Xray policy to `ml-models-local`
Share model across 5 teams without re-downloading	All teams pull from JFrog — cached once

Next Steps¶

👉 MLOps Pipeline with JFrog 👉 AI/ML Security with Xray

🧠 Quick Quiz¶

#

What JFrog feature lets you search for the current production ML model without knowing its exact filename?

📬 DevopsPilot Weekly — Learn DevOps, Cloud & Gen AI the simple way.
👉 Subscribe here