Download Model Weights

To run your Roboflow models on your own hardware, you can either use Roboflow Inference (the recommended, automatic method) or manually download Model Weights (for specific edge cases).

Automatic Download via Roboflow Inference (Recommended)

Roboflow Inference is our open-source, scalable system for running models locally on CPU and GPU devices.

This is the fastest and most reliable way to get started. When you use Inference, you don’t need to manage files or versioning; Roboflow Inference automatically fetches and caches your model weights the first time you run your code.

• How it works: On your first inference request, the weights are downloaded from Roboflow’s servers and stored locally. All future predictions use this local cache - images are not sent to the cloud.

• Deployment options:

  • Workflows: For production-ready multi-step computer vision workflows

  • Python inference SDK: For Python integration

Manual Model Weights Download

Sometimes you may need the raw weights file (ex: a PyTorch .pt file) to run on devices Roboflow does not yet natively support, such as custom Android implementations. The formats you can download depend on the model architecture.

See the Supported Models table for weights download compatibility.

Core, including the Free Tier, supports downloading weights through the Python SDK. Classification weights require separate workspace access, and SAM3 and RF-DETR NAS weights cannot be downloaded. Download access does not change your model's commercial-license terms.

Export Formats

ONNX is the recommended export format. It is available for the widest range of architectures, and you can convert an ONNX file to run on TensorRT, OpenVINO, CoreML, or any other ONNX-compatible runtime. Roboflow Inference also runs models with ONNX Runtime.

Format	Typical use	Available for
ONNX	Cross-platform runtimes (ex: TensorRT, OpenVINO)	Roboflow 3.0, YOLOv8, YOLOv12, YOLO26, YOLO-NAS, RF-DETR, and classification models
PyTorch (.pt)	Custom PyTorch pipelines	Roboflow 3.0, YOLOv8, YOLOv12, YOLO26, YOLO-NAS, RF-DETR, and classification models
CoreML	iOS and Apple devices	Roboflow 3.0, YOLOv8, YOLOv12, YOLO26, RF-DETR, and classification models
TFLite (int8)	Microcontrollers (ex: OpenMV)	Roboflow 3.0, YOLOv8, YOLOv12, and YOLO26
TensorFlow.js	Web browsers	Roboflow 3.0 and RF-DETR object detection models
Luxonis OAK / DepthAI	OAK cameras	Roboflow 3.0 and YOLOv8 object detection models

• Instance segmentation, classification, and keypoint detection models export to ONNX. CoreML is also available for instance segmentation and classification models.

• RF-DETR models can also export to TFLite, TensorRT, and ExecuTorch with the open source rf-detr library (experimental).

• RF-DETR NAS and SAM3 models run only through Roboflow Inference and do not support weight download.

• Fine-tuned multimodal models (ex: PaliGemma, Florence-2, Qwen) download as their Hugging Face output directory (safetensors) without conversion.

Method A: Roboflow Platform

Navigate to the model within your Project. If your workspace has the "Download Weights" button, click it to download the formats available for your model's architecture. Otherwise, use Method B.

Method B: Python SDK

You can also use the Roboflow Python package to download weights directly to your directory:

from roboflow import Roboflow

rf = Roboflow(api_key="YOUR_API_KEY")
model = rf.workspace().project("PROJECT_ID").version(1).models()[0]
model.download()  # Downloads 'weights.pt' to your local folder

Note: Roboflow does not provide technical support for model weights used outside of the Roboflow Inference ecosystem. For the best experience, we recommend using the Inference path outlined in Section 1.