Roboflow Inference turns a computer or edge device into a server for computer-vision models. It serves fine-tuned models you host yourself alongside foundation models such as Florence-2, CLIP, and SAM2, and exposes them over an HTTP API that defaults to port 9001, documented at /docs and /redoc on the running instance. A Python client package talks to the server, and other languages use the REST API directly.
Workflows are the composition layer: blocks of common functionality that give models a shared interface so they can be chained and swapped. A workflow can detect, classify, and segment, call large multimodal models at any stage, apply traditional computer-vision methods such as OCR, barcode and QR reading, and template matching, then track, count, time, measure, and visualize the results, add business logic, and push output to external systems through email, SMS, and webhook blocks. That makes a workflow either a vision micro-service inside a larger application or a self-contained visual agent running on a video stream.
Video is handled by the server rather than by the caller. It runs workflows against RTSP streams and webcam devices and manages hardware acceleration, multiprocessing, decoding, and GPU batching; pipeline results are consumed over the API for downstream processing. The server picks a container image suited to the host, using CUDA acceleration where a compatible GPU is present, and runs on Linux, Windows, macOS, NVIDIA Jetson, Raspberry Pi, and cloud machines.
Some capabilities are open and others need an account key, with metered usage:
| Capability | Open access | With API key (metered) |
|---|---|---|
| Pre-trained and foundation models | yes | yes |
| Video stream management | yes | yes |
| Dynamic Python blocks | yes | yes |
| Public workflows | yes | yes |
| Private workflows | — | yes |
| Fine-tuned and Universe models | — | yes |
| Active learning | — | yes |
| Serverless hosted API | — | yes |
| Dedicated deployments | — | yes |
| Device management and model monitoring | — | enterprise |
Features
- Self-hosted models: run your own fine-tuned models on your own hardware
- Foundation models: Florence-2, CLIP, and SAM2 among the pre-trained models reachable without a key
- Workflows: composable blocks chain and swap models, with a visual editor and an API for running a workflow by workspace and ID
- Video pipelines: RTSP streams and webcam devices with hardware acceleration, multiprocessing, decoding, and GPU batching
- Notifications and sinks: email, Twilio SMS, and webhook blocks push results into external systems
- Classical computer vision: OCR, barcode and QR reading, and template matching alongside model inference
- APIs and client SDK: a REST API with OpenAPI and Redoc documentation, plus the inference_sdk Python client
- Custom blocks: dynamic Python blocks and your own models extend the block library
- Hardware range: Linux, Windows, macOS, NVIDIA Jetson, Raspberry Pi, and cloud servers, with automatic container selection
- Hosted options: dedicated deployments billed hourly and a serverless hosted API billed per call
- Example gallery: published workflows for small-object detection with SAHI, multi-model consensus, active learning, licence-plate reading, face blurring, and background removal
