OpenCV 5 ships with a rebuilt DNN engine that covers over 80% of ONNX operators, up from 22% in 4.x. RISC-V hardware acceleration is first-class for the first time. LLMs and VLMs run through the same Net API used for YOLO. For embedded teams building vision products with AI inference, the runtime choice (OpenCV, ONNX Runtime, vendor SDK, or custom path) just collapsed into one.
The new DNN engine is a typed operation graph with shape inference, constant folding, and operator fusion. Dynamic shapes work. On a standard x86 benchmark, OpenCV 5 DNN runs XFeat 31% faster than ONNX Runtime and YOLOv8n 11.5% faster. The RISC-V path is the more significant change for hardware builders: embedded targets with RISC-V AI accelerator blocks can now run the same inference pipeline a developer debugs on x86, with no separate runtime and no binary blob dependency. The ENGINE_AUTO mode tries the new engine first and falls back to the classic engine, so migrating from OpenCV 4.x carries minimal integration risk.
The vendors most exposed are those selling proprietary embedded inference SDKs that competed on ONNX coverage. OpenCV 5 at 80% is not complete, but it clears the threshold where most production vision workloads run. Teams that avoided RISC-V AI hardware because the inference stack was fragmented have a production-ready path now. The OpenCV project is an Arm Gold member; the RISC-V and Arm acceleration paths in version 5 are the result of sustained sponsorship, not a volunteer weekend project.