Tencent's Hunyuan team has launched Large-Vision, a 52B-parameter multimodal AI model featuring MoE architecture. The model supports any resolution images, videos, and 3D inputs without preprocessing, offering enhanced multilingual understanding for global applications. This advancement positions Tencent competitively in the multimodal AI landscape.
Salesforce and the University of Southern California have developed CoAct-1, a hybrid AI agent system combining code and GUI operations. This innovation addresses the fragility of traditional GUI agents, achieving a 60.76% success rate on benchmark tests while reducing task steps by 33%. The technology promises enterprise applications in customer support, sales, and marketing automation.
Zhipu has released GLM-4.5V, a groundbreaking open-source visual reasoning model with 106B parameters. Achieving state-of-the-art performance in 41 multimodal benchmarks, it excels in image, video, and document understanding. The model is now available on ModelScope and Hugging Face, offering developers advanced AI capabilities for real-world applications.
Zhipu Technology has launched GLM-4.5V, a cutting-edge visual understanding model with 106 billion parameters. The model features a 'thinking mode' switch and excels in tasks like distinguishing between McDonald's and KFC fried chicken. Available on open-source platforms, it also supports real-time screenshot analysis for practical applications.
Baidu AI has launched PaddleOCR 3.1, featuring major upgrades in multilingual recognition (37 languages), document translation pipelines, and Model Context Protocol (MCP) server integration. The update delivers 30%+ accuracy improvements and enterprise-grade translation tools for technical documents.
Hong Kong Polytechnic University and OPPO Research Institute have open-sourced DLoRAL, a revolutionary dual-LoRA framework for video super-resolution. The technology enhances spatial details while maintaining temporal consistency, offering 10x faster processing than traditional methods. Available on GitHub, it marks a significant leap for RealVSR applications.
