On March 5, 2026, the YuanLab.ai team officially released the open-source version of Yuan3.0 Ultra, a multimodal foundational large model. As the flagship model in the Yuan 3.0 series designed for trillion-parameter scale, Yuan3.0 Ultra is one of only three trillion-parameter open-source multimodal large models in the industry today. The model systematically introduces training efficiency optimizations for Mixture of Experts (MoE) large models into its structural design and is deeply optimized for enterprise applications and intelligent agent tool invocation. Yuan3.0 Ultra adopts a unified multimodal model architecture, consisting of a visual encoder, a language backbone network, and a multimodal alignment module, enabling collaborative modeling of visual and language information. The language backbone network is built on an MoE architecture, comprising 103 Transformer layers. At the initial training stage, the model has 1515 billion parameters, which are optimized to 1010 billion parameters using the LAEP method, improving pre-training computational efficiency by 49%, with 68.8 billion active parameters. Additionally, the model introduces a Localized Filtering Attention (LFA) mechanism, effectively enhancing its ability to model semantic relationships and improving model accuracy. Yuan3.0 Ultra excels in enterprise-level tasks such as multimodal document understanding, retrieval-augmented generation (RAG), tabular data analysis, content summarization, and tool invocation. It can handle complex information formats in enterprise environments with high quality, such as mixed text-image documents, multi-level structured tables, and cross-document knowledge retrieval, providing core capability support for building multimodal data-driven enterprise Agent AI based on intelligent agent frameworks like OpenClaw. Currently, Yuan3.0 Ultra is fully open-source, with both model parameters and code available for free download and use.