mnmly/rfdetr-large-2026-mlx-fp32

RF-DETR Large (2026) — MLX, fp32

MLX-format conversion of Roboflow's RF-DETR Large (2026) object-detection model, for use with mlx-swift-rf-detr on Apple Silicon.

• Upstream: roboflow/rf-detr (Apache-2.0), checkpoint rf-detr-large-2026.pth
• Task: object detection — COCO, 90 classes
• Resolution: 704 · decoder layers: 4 · queries: 300 · backbone: dinov2_windowed_small (hidden 256)
• Precision: fp32
• Files: config.json + preprocessor_config.json + model.safetensors

Conversion

Converted from the upstream PyTorch checkpoint with Scripts/convert_detection.py large-2026. Config values are read from the upstream RFDETRLargeConfig; weights keep their model. prefix and NCHW conv layout (the Swift loader transposes conv weights to NHWC at load).

Usage (Swift)

import MLXRFDETR

let dir = URL(fileURLWithPath: "rfdetr-large-2026-mlx-fp32")   // this repo's downloaded files
let predictor = try MLXRFDETR.fromPretrained(dir, scoreThreshold: 0.3, nmsThreshold: 0.5)
let result = try predictor.predict(url: URL(fileURLWithPath: "image.jpg"))

License

Apache 2.0 — see LICENSE. Model architecture and weights © Roboflow, Inc. This repository redistributes a format-converted copy of the upstream Apache-2.0 weights; no weights were retrained or modified beyond the PyTorch→MLX serialization.