Roboflow Inference

An open source inference server from Roboflow, Inference turns a computer or edge device into a deployment target for vision models, from fine-tuned YOLOv5 through YOLOv12 detectors to foundation models such as CLIP, SAM 2, and Florence-2. Models run behind an HTTP API started with a single Docker command or a pip install of inference-cli, with automatic CUDA acceleration when a GPU is present and CPU fallback otherwise, and prebuilt images cover x86 CPU and GPU, TensorRT, AWS Lambda, and four NVIDIA Jetson JetPack generations. A Workflows system chains models, logic, and external API calls into vision pipelines, and the server handles live video streams as well as still images. The core carries an Apache 2.0 LICENSE.core covering everything except bundled models with their own terms and the enterprise directories held under a separate Roboflow Enterprise License. The same engine powers Roboflow's paid hosted API, so projects move between local and cloud without code changes, while model management ties into the platform through an API key. At roughly 2,400 GitHub stars it is a practical route from trained model to endpoint.