Roboflow

Best Real-Time Vision Models

Compare the best 23 real-time vision models and try 7 of them on your own image, free in the Roboflow Playground. 23 are open-weight, so you can self-host them for free under their licenses.

23 models · 23 open-weight · 7 free to try · prices synced Aug 1, 2026

Newest

RF-DETR Segmentation

Most recently added (Oct 2025)

Open-Source Real-Time Vision Models

23 models with downloadable weights you can self-host under their licenses (Apache 2.0, AGPL 3.0, and GPL v3). 7 run live in the Playground through hosted APIs, so self-hosting is optional.

RF-DETR SegmentationRoboflow
33.6M-38.6MApache 2.0Oct 2025
Try
RF-DETRRoboflow
30.5M-126.9MApache 2.0Mar 2025
Try
YOLOETHU-MIG
10M-50MAGPL 3.0Mar 2025
YOLOv12THU-MIG
2.6M-59.1MAGPL 3.0Feb 2025
DEIMIntellindust AI Lab
4M-62MApache 2.0Dec 2024
D-FINEUSTC
4M-62MApache 2.0Oct 2024
YOLO11Ultralytics
2.6M-56.9MAGPL 3.0Sep 2024
Try
Tencent
YOLO WorldTencent AI Lab
13MGPL v3Feb 2024
Try

How to Choose a Real-Time Vision Model

Real-time choices are budget decisions: a frame arrives every 33 milliseconds at 30 FPS, and everything (inference, pre and post processing, tracking) must fit inside it on the hardware you actually deploy.

The same architecture ships in sizes from nano to extra-large, and the right size is set by your hardware and frame rate, not by peak benchmark accuracy. A model that hits 80 FPS on a data-center GPU may run at 4 FPS on the edge box you can afford to put on site. Fix the target device and required FPS first, then choose the largest model that fits with headroom; published FPS numbers are upper bounds measured on strong GPUs.

Input resolution is the strongest lever in the family: halving it can double throughput, at the price of missing small objects. Detection-plus-tracking often beats detecting on every frame (detect every few frames, track between them). And not every pipeline needs real time at all; if the video is recorded and analyzed later, batch processing with a bigger, more accurate model is usually the better trade.

A compact model fine-tuned on your own footage routinely beats a larger generic one on both accuracy and speed for a fixed camera and class list. Before stepping up a size (and losing your latency headroom), try training the size you have on data from the actual deployment.

The bottom line: Fix the device and frame rate first, size the model to fit with headroom, and fine-tune small on your own footage before reaching for a bigger architecture.

Frequently Asked Questions About Real-Time Vision Models

Real-time vision models are architectures engineered to keep up with live video, processing frames in the milliseconds-per-frame range on practical hardware rather than seconds per image. The family is dominated by fast detector architectures (RF-DETR, the YOLO line, RT-DETR) plus latency-optimized variants for segmentation and pose. The speed comes from architectural budget choices: compact backbones, single-pass prediction, and resolutions tuned to the deployment target, trading a little peak accuracy for the ability to run on every frame, often on edge devices without a data-center GPU. This page lists 23 real-time vision models, including 23 open-weight options you can self-host; 7 of them run live in the Playground so you can test them on your own images.

The working threshold is keeping up with the camera: about 30 frames per second, or roughly 33 ms per frame end to end. Whether a given model hits that depends on hardware, input resolution, and batch size as much as architecture, so treat published FPS numbers as upper bounds measured on strong GPUs and validate on the device you will actually deploy to. A VLM answering in one to several seconds is fast for a language model and still two orders of magnitude away from real-time video.

It depends on your task and constraints. For fixed categories in production, a model fine-tuned on your own data typically beats any general-purpose model. Compare the real-time vision models on this page and try them on your own image to see which fits.

Yes. 23 of the 23 models here are open-weight (for example RF-DETR Segmentation, RF-DETR, and YOLOE), free to self-host under their licenses (Apache 2.0, AGPL 3.0, and GPL v3).

Yes. You can run 7 of them in the Roboflow Playground for free. Upload an image and compare the models' output side by side, no setup required.

This page lists all 23 real-time vision models in the Roboflow Playground catalog, all of them open-weight and free to self-host; 7 of them run live in the Roboflow Playground on your own images. Compare licenses, parameters, and release dates side by side, or open any model page for full details.