Darknet refuses to die, and now it chases 1,000 FPS

What it does

Darknet is the original open-source neural network framework that powers YOLO object detection. Written in C, C++, and CUDA, it trains and runs inference on Linux, Windows, and Mac, with support for NVIDIA CUDA and AMD ROCm GPUs. The project provides pre-trained weights for standard datasets like MSCOCO and niche ones like LEGO Gears, and claims to process LEGO dataset videos at up to 1000 FPS on an NVIDIA RTX 3090.

The interesting bit

This is not a rewrite; it is a custody battle won by discipline. After Joseph Redmon and Alexey Bochkovskiy’s branches went dormant, Hank.ai took over in 2023 and has been systematically modernizing the C++17 codebase—unifying CMake builds, replacing legacy C APIs, and stripping out unmaintained commands—while insisting the framework is still faster than newer YOLO incarnations. The repo itself is now a mirror from Codeberg, where active development happens.

Key highlights

• Claims 1000 FPS inference on LEGO dataset videos using an RTX 3090 (Darknet v3 “Jazz”), with further optimizations in v4 and v5.
• Cross-platform CMake build for Linux, Windows, and Mac; CPU-only builds work on Raspberry Pi, GPU builds require NVIDIA CUDA or AMD ROCm.
• New C and C++ APIs introduced in v3 to replace the legacy interface, plus sample applications in src-examples.
• Experimental ONNX export added in v5, alongside Profile-Guided Optimization and OpenBLAS support for CPU builds.
• Free for commercial use with no licensing fees.

Caveats

• The GitHub repository is now a mirror from Codeberg as of August 2025, so primary development and issues have moved.
• ONNX export and Java language bindings are marked experimental and incomplete, respectively.
• A prediction structure change in v5.1 means any application linking against darknet.h or darknet.hpp must be recompiled to avoid segfaults.
• The legacy C API was modified in v3, so older integrations need minor updates.

Verdict

Worth a look if you need a bare-metal, C++-first object detection pipeline and prefer native builds over Python-heavy ecosystems. Skip it if you want turnkey Python notebooks or a project with a quiet GitHub presence.