Roboflow Inference provides self-hosted computer vision inference servers and workflows for edge and on-prem deployments, with Dockerized components and API access. It emphasizes running models, managing video streams, and integrating with Workflows.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Inference is a Python-based tool to turn any computer or edge device into a command center for computer vision projects. It supports self-hosted deployment and interaction via an API, with features for workflows, video processing, and external integrations.
How it works
The project offers an API and SDK (inference-sdk) to run models and workflows on images and video streams. It includes components for managing video pipelines, WebRTC streaming, and integration blocks within Workflows. It can load models, run workflows, and provide endpoints for REST/OpenAPI access. Notable capabilities shown in the README include opt-in pre-loading of models for workflows, offline/air-gapped deployments, and video processing with GPU acceleration where available.
Getting started
Quickstart instructions in the README indicate:
pip install inference-cli && inference server start --dev
This starts the inference server in development mode and provides a local notebook at http://localhost:9001/notebook/start. The README also provides a Python example to connect via the SDK:
pip install inference-sdk
from inference_sdk import InferenceHTTPClient
client = InferenceHTTPClient(
api_url="http://localhost:9001", # use local inference server
# api_key="<YOUR API KEY>" # optional to access your private data and models
)
result = client.run_workflow(
workspace_name="roboflow-docs",
workflow_id="model-comparison",
images={
"image": "https://media.roboflow.com/workflows/examples/bleachers.jpg"
},
parameters={
"model1": "yolov8n-640",
"model2": "yolov11n-640"
}
)
print(result)
For connecting to video streams, there is an example showing how to start an inference pipeline with a workflow and consume results via the API.
Recent releases
Latest releases (as of the README) include:
- v1.3.8 (2026-07-31): Opt-in model pre-loading for Workflows in InferencePipeline
- v1.3.7 (2026-07-27): OFFLINE_MODE — air-gapped deployments
- v1.3.6 (2026-07-22): NVIDIA Cosmos 3 Edge — initial preview
- v1.3.5 (2026-07-10): NVIDIA Cosmos 3 Edge preview
- v1.3.4 (2026-07-06): Several fixes and changes to HTTPS gateway handling
Traction
Stars: 2400. Forks: 295. Open issues: 154. Language: Python. License: none listed.
Behind the repo
Not applicable from the provided content.
Caveats
License is listed as none; core license is Apache 2.0 per the docs, but the README section also notes license details for models and cloud features. The README includes references to multiple platforms (Linux, Windows, Mac, NVIDIA Jetson, Raspberry Pi) and various deployment modes (on-prem, cloud, edge).