Roboflow Inference

Roboflow Inference turns a computer or edge device into a computer vision server. It runs your own fine-tuned models and foundation models such as Florence-2, CLIP, and SAM2 on images and video streams, and exposes them through a REST API and a Python SDK.

Workflows are its main building block: composable blocks with a common interface that detect, classify, and segment objects, call large multimodal models at any stage, track, count, time, and measure objects, and combine models with traditional methods such as OCR, barcode and QR reading, and template matching. A workflow can serve as a vision microservice inside a larger application or as a self-contained visual agent that runs on a video stream. Example workflows cover small-object detection with SAHI, multi-model consensus, active learning, license plate reading, face blurring, and background removal.

The server runs workflows on RTSP streams and webcams, handling hardware acceleration, video decoding, multiprocessing, and GPU batching. Pipeline outputs are consumed through the API or sent onward by notification blocks such as email, Twilio SMS, or a webhook.

Inference runs self-hosted on Linux, Windows, Mac, NVIDIA Jetson, Raspberry Pi, or your own cloud, with automatic selection of a GPU-accelerated container where CUDA is available. Pre-trained models, foundation models, video stream management, custom Python blocks, and public workflows work without an account. A Roboflow API key adds private workflows, fine-tuned and Universe models, active learning, a serverless hosted API, and dedicated deployments.