On June 8, the open-source computer-vision library OpenCV released version 5.0 on the opening day of CVPR 2026. The DNN inference engine has been completely rewritten: the old engine that covered only ~22% of ONNX operators has been replaced by a typed-graph-based new engine, with coverage jumping to over 80%, and operators for dynamic shape, If/Loop subgraphs, constant folding and QDQ, BatchNorm, Attention fusion, etc., are all added.
More notable is that the new engine brings LLM and VLM into the DNN module for the first time: built-in tokenizer, attention, and KV-cache enable Qwen 2.5, Gemma 3, PaliGemma, and GPT-2 family models to share the same Net API with YOLO. On an Intel Core i9-14900KS, against ONNX Runtime, XFeat is 31% faster, BiRefNet 32.4% faster, and OWLv2 36.6% faster.
Limits are equally clear: the new engine currently supports CPU only, CUDA/OpenVINO users still need the classic engine or ONNX Runtime; C++17 becomes the minimum standard, and the Caffe/Darknet parsers and legacy C API are retired.
OpenCV 5.0's real value isn't benchmark refresh, but unifying "classic vision algorithms + modern multimodal models" onto the same runtime. Previously you had to stitch together OpenCV + ONNX Runtime + Transformers to build a visual-QA or image-captioning pipeline; now a single cv::dnn::readNet call can run it all. For industrial inspection, robotics, AR/VR and other edge scenarios, "one fewer dependency" is often more meaningful than a few percentage points of throughput.