1.4.0

roboflow/rf-detr1.4.0Jan 22, 2026by SkalskiP

AI Summary

A major feature release introducing new large model sizes, support for YOLO format datasets, and dropping Python 3.9 support.

Key Highlights

Dropped support for Python 3.9, requiring Python 3.10 or newer.
Added new pre-trained checkpoints: Object Detection (L, XL, 2XL) and Instance Segmentation (N, S, M, L, XL, 2XL).
Enabled training on datasets in YOLO format.
Simplified project dependencies by removing several packages from `pyproject.toml`.
Fixed precision, recall, and F1 computation during confidence sweeps.

Breaking Changes

Dropped support for Python 3.9

New Features

New pre-trained checkpoints (L, XL, 2XL for detection; N, S, M, L, XL, 2XL for segmentation).
Support for training on YOLO format datasets.

Full Release Notes

# Changelog

> [!WARNING]
> Starting with version `1.4.0`, RF-DETR drops support for Python `3.9`. If your environment still relies on Python `3.9`, stay on RF-DETR `1.3.x` or upgrade your Python runtime to `3.10` or newer.

## 🚀 Added

- New pre-trained checkpoints. Object detection includes new L, XL, and 2XL checkpoints. Instance segmentation includes N, S, M, L, XL, and 2XL checkpoints. ([#539](https://github.com/roboflow/rf-detr/pull/539))

    <img alt="rf_detr_1-4_latency_accuracy_object_detection" src="https://storage.googleapis.com/com-roboflow-marketing/rf-detr/rf_detr_1-4_latency_accuracy_object_detection.png" />
    
    <img alt="rf_detr_1-4_latency_accuracy_instance_segmentation" src="https://storage.googleapis.com/com-roboflow-marketing/rf-detr/rf_detr_1-4_latency_accuracy_instance_segmentation.png" />
    
    ```python
    import requests
    import supervision as sv
    from PIL import Image
    from rfdetr import RFDETRSegMedium
    from rfdetr.util.coco_classes import COCO_CLASSES
    
    model = RFDETRSegMedium()
    
    image = Image.open(requests.get('https://media.roboflow.com/dog.jpg', stream=True).raw)
    detections = model.predict(image, threshold=0.5)
    
    labels = [
        f"{COCO_CLASSES[class_id]}"
        for class_id
        in detections.class_id
    ]
    
    annotated_image = sv.MaskAnnotator().annotate(image, detections)
    annotated_image = sv.LabelAnnotator().annotate(annotated_image, detections, labels)
    ```

    https://github.com/user-attachments/assets/add23fd1-266f-4538-8809-d7dd5767e8e6

- Support for training object detection and instance segmentation models using datasets in YOLO format. ([#569](https://github.com/roboflow/rf-detr/pull/569))

## 🌱 Changed

- Simplified project dependencies by removing `cython`, `fairscale`, `timm`, `accelerate`, `ninja`, `einops`, `pandas`, `pylabel`, and `open_clip_torch` from `pyproject.toml`. This reduces the dependency footprint and makes RF-DETR easier to install alongside other Python packages. ([#571](https://github.com/roboflow/rf-detr/pull/571))

## 🔧 Fixed

- Fixed precision, recall, and F1 computation during confidence sweeps. This resolves an issue where recall values were identical across classes and aligns per-class and class-averaged metrics with expected COCO-style behavior. ([#545](https://github.com/roboflow/rf-detr/pull/545))

# 🏆 Contributors

@isaacrob ([Isaac Robinson](https://www.linkedin.com/in/robinsonish/)), @probicheaux ([Peter Robicheaux](https://www.linkedin.com/in/peter-robicheaux-01958813b/)), @Matvezy ([Matvei Popov](https://www.linkedin.com/in/matvezy/)), @mkaic ([Kai Christensen](https://www.linkedin.com/in/mkaic/)), @anujonthemove ([Anuj Khandelwal](https://www.linkedin.com/in/anujonthemove/)), @brunopicinin ([Bruno Cardoso](https://www.linkedin.com/in/brunopicinin/)), @capjamesg ([James Gallagher](https://www.linkedin.com/in/jg12927/)), @Borda ([Jirka Borovec](https://www.linkedin.com/in/jirka-borovec/)), @SkalskiP ([Piotr Skalski](https://www.linkedin.com/in/skalskip92/))