Roboflow Inference

What it does

Roboflow Inference is a self-hosted server that runs computer vision models and Workflows on images, RTSP streams, and webcam feeds. It takes in images, video streams, and model or workflow requests over a local API, and returns predictions, workflow outputs, and pipeline results for downstream processing.

People use it to run their own fine-tuned models or foundation models such as Florence-2, CLIP, and SAM2 on local hardware. Workflows let users chain models with traditional methods like OCR, barcode reading, and template matching, then track, count, time, and visualize objects. Pipeline outputs can trigger notifications through Email, Twilio, or a Webhook block.

Roboflow Inference suits developers building vision applications on their own machines or edge devices. The server can be installed through the documented container workflow, with NVIDIA Container Toolkit required for supported CUDA acceleration. Running images and video on your own hardware does not consume Roboflow cloud credits. You still pay for the machine and must check both server and model licenses, particularly for commercial use. Hosted services have separate platform pricing.

What it can do

• Run self-hosted fine-tuned computer vision models
• Run foundation models such as Florence-2, CLIP, and SAM2
• Build Workflows that chain models with OCR, barcode, QR, and template matching
• Track, count, time, measure, and visualize detected objects
• Process RTSP streams, webcams, and video with GPU batching
• Send notifications or call external APIs through Workflow blocks

Works with

• Email notification block
• Twilio notification block
• Webhook block
• Telegram bot (in tutorial example)

Example workflows

Live stream text-match detection

Illustrative example. Input: RTSP stream and a text prompt. Steps: start a pipeline with the clip-frames Workflow, consume results via the SDK, compare similarity to a threshold. Output: per-frame match flag and similarity score.

Model comparison on an image

Illustrative example. Input: an image URL and two model names. Steps: call run_workflow with the model-comparison Workflow via the SDK. Output: the Workflow result comparing the two models' predictions.

Before you choose

• Provide suitable compute, memory and drivers for the selected model.
• Check server and model licenses separately before commercial deployment.
• Hosted API and cloud workflow charges belong to the separate Roboflow platform.

Listed under

• Coding
• Sandboxes and infrastructure
• Computer vision
• Computer-vision
• Developers
• Edge devices
• Edge-devices
• Object detection

Get started

Run pip install inference-cli and inference server start --dev.

1. Install Docker and, for GPU use, the NVIDIA Container Toolkit.
2. Run pip install inference-cli && inference server start --dev to pull the image and start the server.
3. Open the notebook quickstart at http://localhost:9001/notebook/start and install inference-sdk to call the API.

Processing images and video on your own hardware does not consume Roboflow cloud credits. You supply the compute. Review the server/component licenses and commercial rights for each model; some models require additional licensing. Roboflow hosted services use separate platform plans.

Roboflow Inference is free to use. See their pricing page ↗

Sources for this overview

Product pages, documentation, and pricing checked 29 Oct 2026. Each source supports part of the overview; availability and prices can change.

• GitHub - roboflow/inference: Turn any computer or edge device into a command center for your computer vision projects. · GitHub ↗github.com
• Roboflow Inference - Roboflow Inference ↗inference.roboflow.com
• Roboflow Pricing: Free, Core & Enterprise Plans ↗roboflow.com

Company website

inference.roboflow.com: product, pricing and policy pages

Hacker News

Launch posts and discussion

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 29 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.