Overview

Roboflow covers the whole path from raw images to a model running on a device. You upload footage, annotate it with model assistance, train either one of its own architectures or a third-party model like YOLO26, then deploy to the hosted API, a VPC, or a Docker container sitting on a Jetson inside the factory network. Workflows is the piece that changed the product: a visual editor where you chain a detector into a classifier into an OCR block into custom Python, version the whole thing and ship it as one deployable unit.

The company also does real research. RF-DETR pairs a DINOv2 pretrained encoder with a multiscale DETR head, was accepted at ICLR 2026, and reports 56.5 AP on COCO at 6.8 ms on a T4 under TensorRT FP16, with a 2XL variant clearing 60 AP. That is the first real-time model past 60 on that benchmark. Roboflow has raised roughly $63.6M across four rounds and had about 137 staff as of June 2026, so it is a going concern rather than a side project that might vanish mid-deployment.

Best fit is visual inspection, counting, safety monitoring or anything where the data is proprietary and inference has to happen on site. Core at $79 a month covers a small private project with three seats. Anything approaching a fleet of edge devices with commercial inference rights lands you in a custom Enterprise quote, and that should be settled early. Skip it if you are training foundation-scale models or need full control of the training loop, because the abstraction that makes it quick also hides knobs you will eventually want.

Key Features

✓ RF-DETR, the in-house detection architecture accepted at ICLR 2026, reports 56.5 AP on COCO at 6.8 ms on an NVIDIA T4 under TensorRT FP16, with a 2XL variant past 60 AP

✓ Workflows visual builder chains detection, classification, OCR, SAM 3 segmentation and custom Python blocks into one versioned, deployable pipeline

✓ Self-hosted Inference server ships as a Docker container tuned for NVIDIA Jetson on JetPack 6.2 and runs fully air-gapped with cached weights

✓ Trains and hosts third-party architectures as well as its own, including YOLO26, YOLO11, YOLOv8 and YOLOv5

✓ AI-assisted annotation using foundation models to pre-label before human review

✓ Universe library of open datasets and pretrained models for cold-starting a class you have no data for

My Take

RF-DETR now posts 56.5 AP on COCO at 6.8 ms on a T4, which puts an openly licensed detector in the same conversation as the AGPL YOLO line. For a factory deployment that licence difference matters more than another point of mAP. The labelling tool people remember is now the smaller half of the product. Where it gets expensive is the jump from Core at $79 a month to a custom Enterprise quote, which is the gate for commercial inference rights on your own edge boxes, and that conversation should happen before the pipeline is built rather than after.