Changelog

You can now multiple annotations in Roboflow Annotate. This is useful for repositioning or updating multiple annotations at once.

RF-DETR is our new, state-of-the-art, real-time object detection model. The model architecture is open source and also available for training and deployment with Roboflow Inference.

You can now merge detections into a single bounding box with the Detections Merge Workflows block.

You can now pass RTSP streams as an input to a Workflow in the Workflows web editor for use in testing your Workflow.

Batch Processing lets you upload batches of images and videos and run Roboflow Workflows on the data. You can then download the results for futher processing.

Inference 0.40.0 adds a Camera Calibration block for use in calibrating cameras for vision systems and speed ups to PyTorch models.

You can now train YOLOv12 models on Roboflow and deploy them using the Roboflow Serverless Hosted API.

You can now use Qwen2.5-VL with a Dedicated Deployment in Roboflow Workflows, and deploy the Workflow on your own hardware or with your Dedicated Deployment.

The Roboflow Workflows web editor now has an Ask AI button you can use as a partner in building your Workflows.

You can now train a computer vision model in minutes with Roboflow Instant.