// the find
roboflow/inference
Turn any computer or edge device into a command center for your computer vision projects.
Roboflow Inference is a self-hostable CV inference server that runs on anything from a Raspberry Pi to an NVIDIA Jetson to a cloud GPU. It wraps YOLO variants, SAM2, Florence-2, CLIP, and a dozen other models behind a unified REST API, plus a 'Workflows' DAG system for chaining CV operations without writing glue code. The target user is someone building a production CV pipeline who doesn't want to hand-roll model serving, batching, and stream management.
The Workflows abstraction is the real differentiator — composable blocks for tracking, counting, OCR, barcode reading, and webhook sinks mean you can build a license plate reader or dwell-time monitor without any custom code. Edge device support is genuinely broad: separate Dockerfiles and JetPack-specific builds for Jetson 5.1.1 through 7.1.0, Raspberry Pi, and TensorRT, not just a generic ARM image. The video pipeline handling (RTSP, hardware decoding, GPU batching, multiprocessing) is production-grade and would take weeks to build from scratch. CI is thorough — separate GPU/CPU integration test workflows, Jetson device tests, Colab tests, and E2E hosted tests.
The licensing story is a mess you'll hit after the honeymoon: the core is Apache 2.0, but enterprise features are source-available under a separate license, model licenses vary per architecture, and cloud-connected features require a Roboflow account with metered billing — none of this is obvious until you're already invested. The Workflows UI lives on app.roboflow.com, not self-hosted, so 'fully self-hosted' is only true if you're comfortable building workflows via JSON API rather than the drag-and-drop editor. Private workflows, fine-tuned model weights, and active learning all require an API key, which means your self-hosted server phones home anyway. The Python package split (inference, inference-cli, inference-sdk, inference-gpu, inference-cpu as separate PyPI packages) adds unnecessary friction for anyone not following the Docker path.