Rethinking speed: modular detection that compiles itself faster.
Quick Start • Documentation • Why Switch? • Architecture • Benchmarks
YOLOv5 was revolutionary. But vision has evolved. Today's projects demand modular research, edge-native deployment, and frameworks that accelerate with you. NEXUS isn't an update—it's a re-architecture for the PyTorch 2.x era.
40% fewer dependencies. 2-3x faster edge inference. 100% pluggable.
| Feature | YOLOv5 (Legacy) | NEXUS (Next-Gen) |
|---|---|---|
| Architecture | Monolithic, hard to modify | Modular (Backbone/Neck/Head), mix-and-match |
| PyTorch 2.x | Partial support | Native torch.compile & sdpa |
| ONNX Export | Static shapes | Dynamic axes + built-in quantization |
| Dependencies | Heavy (~45 packages) | 40% leaner, conflict-free |
| Research Speed | Fork & modify core | Plug new components, keep the core |
| Edge Inference | Baseline | 2-3x faster with optimized ONNX |
| Customization | Edit YAML configs | Pythonic component API |
# Install from PyPI (recommended)
pip install nexus-cv
# Or install from source for latest features
git clone https://github.com/sovereign-ai/nexus.git
cd nexus
pip install -e .import nexus as nx
# Load a pre-trained model (auto-downloads)
model = nx.load("nexus-m") # nano, small, medium, large, xl
# Run inference on an image
results = model.predict("https://ultralytics.com/images/bus.jpg")
# Show results with bounding boxes
results.show()
# Export to optimized ONNX for edge deployment
model.export(format="onnx", dynamic=True, quantize=True)from nexus.components import backbones, necks, heads
# Build a custom detector in 3 lines
backbone = backbones.EfficientNetV2(pretrained=True)
neck = necks.PANet(channels=[24, 48, 64, 128])
head = heads.YOLOHead(num_classes=80)
model = nx.Model(backbone, neck, head)
model.train(data="coco128.yaml", epochs=100)Input Image
↓
┌─────────────────────────────────────────────────────────┐
│ NEXUS CORE ENGINE │
├─────────────────────────────────────────────────────────┤
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ │
│ │ Backbone │ → │ Neck │ → │ Head │ → Detections │
│ └─────────┘ └─────────┘ └─────────┘ │
│ ↑ ↑ ↑ │
│ [EfficientNet] [PANet] [YOLOHead] │
│ [ResNet] [BiFPN] [RetinaHead] │
│ [SwinT] [NASFPN] [CustomHead] │
└─────────────────────────────────────────────────────────┘
Every component is interchangeable. Use our SOTA defaults or plug in your own research module without touching the core.
Tested on NVIDIA RTX 4090, batch size 32, FP16
| Model | mAP@50 | Latency (ms) | ONNX Edge (ms) | PyTorch 2.x Speedup |
|---|---|---|---|---|
| YOLOv5m | 45.2% | 8.1 | 12.4 | 1.0x |
| NEXUS-m | 45.8% | 5.9 | 4.8 | 1.4x |
| YOLOv5x | 50.1% | 13.2 | 22.1 | 1.0x |
| NEXUS-x | 50.9% | 9.8 | 8.7 | 1.5x |
ONNX Edge inference measured on NVIDIA Jetson Orin (INT8 quantized)
We love YOLOv5. That's why we made switching trivial.
# Your existing YOLOv5 code
import torch
model = torch.hub.load('ultralytics/yolov5', 'yolov5s')
# NEXUS equivalent - same API, more power
import nexus as nx
model = nx.load("yolov5s") # Loads YOLOv5 weights automatically100% backward compatible with YOLOv5 weights. Your trained models work immediately.
- Getting Started - Your first detection in 5 minutes
- Component Zoo - All available backbones, necks, heads
- ONNX Deployment Guide - From PyTorch to edge in one command
- Research with NEXUS - Plug in your novel architecture
- API Reference - Every function documented
- GitHub Issues - For bugs and feature requests
- Discord - Join 5,000+ researchers and engineers
- Weekly Office Hours - Live Q&A with core maintainers
- Paper Club - Discuss latest vision papers, implement together
NEXUS is released under AGPL-3.0. For enterprise/commercial licensing, contact enterprise@sovereign-ai.com.
Built with ❤️ by the SOVEREIGN AI Collective
"Vision shouldn't be a black box. It should be a toolkit."
Star ⭐ this repo if you believe vision should be modular, fast, and open.
The more stars, the more components we add to the zoo.