PaddleOCR has launched PP-OCRv6, a new generation of lightweight, multilingual optical character recognition models now available on the Hugging Face Hub. The release introduces a unified model family spanning three tiers from 1.5 million to 34.5 million parameters, engineered to balance computational efficiency with high accuracy across edge, mobile, and server deployments. Architecturally, all tiers utilize the PPLCNetV4 backbone to ensure consistency. Text detection is upgraded through the RepLKFPN module, a lightweight feature pyramid network optimized for multi-scale recognition of dense, rotated, or low-resolution text. Recognition accuracy is enhanced by the EncoderWithLightSVTR architecture, which merges local context modeling with global attention to handle multilingual characters, screen text, and noisy backgrounds. On official benchmarks, the medium-tier model achieves 86.2 percent detection Hmean and 83.2 percent recognition accuracy, marking a 4.6 percentage point improvement in detection and a 5.1 percentage point gain in recognition over the previous PP-OCRv5 server variant. The small and medium models consolidate support for 50 languages, including Simplified and Traditional Chinese, English, Japanese, and 46 Latin-script languages, eliminating the need for separate regional models. PP-OCRv6 is distributed on Hugging Face in multiple formats, including safetensors and ONNX, and integrates with PaddleOCR 3.7's unified inference engine. Developers can deploy the models using Paddle Inference, Hugging Face Transformers, or ONNX Runtime through standardized API configurations. Designed for practical enterprise and developer workflows, the framework outputs structured text compatible with document parsing, retrieval-augmented generation, and analytics pipelines. An online evaluation environment and updated integration code are now accessible, enabling rapid adoption across constrained local systems and high-throughput industrial OCR applications.