14 lines of Python.
Production-grade CV.
Select your deployment target. Every number below is reproducible in one Colab click.
Top-right quadrant. Every time.
GPU inference at 640px input on a single A10G. Detect dots cluster where speed and accuracy peak simultaneously — the same chart, every dataset.
† All measurements: NVIDIA A10G, CUDA 12.2, PyTorch 2.2, batch size 1. Reproduce →
Same frame. Different finish times.
Processing 120 consecutive COCO val frames. Each bar advances as frames complete. Detect finishes before competitors hit frame 80.
Your stack, already supported.
12 first-class integrations. ONNX to TensorRT to Coral Edge TPU — one export command, zero glue code.
Export any Detect model to ONNX in one line. Deploy to any ONNX-compatible runtime.
FP16/INT8 quantization for NVIDIA GPUs. 2–3× throughput boost over raw PyTorch.
Native Neural Engine inference on M1/M2/M3 Macs and iOS devices.
Optimized inference on Intel CPUs, iGPUs, and VPUs. Ideal for industrial edge.
Android and embedded deployment. Supports delegation to GPU and Edge TPU.
Multi-model, multi-GPU serving with dynamic batching and gRPC/HTTP APIs.
One-click dataset import from Roboflow Universe. 200k+ labeled CV datasets.
Active learning loop: run Detect predictions, correct in Label Studio, retrain.
Automatic experiment tracking, model versioning, and benchmark dashboards.
Nano model runs at 28 FPS on Pi 5. Detect ships a Pi-optimized ONNX export.
INT8 quantized nano model: 290 FPS on a $25 USB Coral accelerator.
All official Detect checkpoints on HF Hub. One-line download with versioning.
The entire API. Not a tutorial excerpt.
This is complete working code — detection, segmentation, pose estimation, and streaming. No boilerplate hidden below the fold.
| 1 | import detect |
| 2 | |
| 3 | # Load a model — nano to large |
| 4 | model = detect.load("small") |
| 5 | |
| 6 | # Run on any source: file, URL, webcam, RTSP |
| 7 | results = model.predict( |
| 8 | source="./birds.mp4", |
| 9 | conf=0.25, |
| 10 | device="cuda", |
| 11 | stream=True |
| 12 | ) |
| 13 | |
| 14 | # Every result: boxes, masks, keypoints, classes |
| 15 | for frame in results: |
| 16 | frame.show() # annotated + timed |