tl;dr: OpenCV 5 ships a rewritten graph-based DNN engine that boosts ONNX operator coverage from ~22% to over 80%, adds dynamic shapes, attention/MatMul fusion, and built-in tokenizer + KV-cache support for running LLMs and VLMs (Qwen, Gemma, GPT) directly via the Net API. Benchmarks show it matching or beating ONNX Runtime on CPU for models like YOLOv8, DINOv2, and OWLv2, while the old engine remains available behind the same API for backward compatibility. The release also modernizes the core with FP16/BF16 types, 0D/1D Mat support, a redesigned hardware acceleration layer (Intel IPP, Arm KleidiCV, Qualcomm FastCV, RISC-V), and split 3D/calibration modules; pip release is slated for June 2026.