Yuan3.0 Ultra - Inspur Information Team's Open Source Multimodal Foundation Model

What is Yuan 3.0 Ultra?

Yuan3.0-Ultra is an open-source, trillion-parameter multimodal foundational model from Inspur's YuanLab.ai team, boasting 1.01T total parameters and 68.8B activation parameters. Employing a MoE architecture, it integrates a visual encoder and a multimodal alignment module, and innovatively proposes the Layer-Adaptive Expert Pruning (LAEP) algorithm, improving pre-training efficiency by 49%. The model addresses the "overthinking" problem through an improved Reflection Suppression Reward Mechanism (RIRM), achieving leading performance in enterprise-level RAG, table understanding, and tool invocation benchmarks such as DocMatix, ChatRAG, and MMTab, positioning it as a core engine for complex enterprise-level agent applications.

Main features of Yuan3.0 Ultra

• Multimodal understandingIt supports joint understanding and reasoning of text, images, and tables, enabling cross-modal information fusion.
• Search Enhancement GenerationIt accurately identifies the content of a company's private knowledge base and provides reliable answers based on evidence.
• Complex document analysisDeeply analyze structured documents such as financial statements and approval forms to extract key information.
• Intelligent summary generationGenerate faithful and concise high-quality summaries based on the source text to ensure accurate and efficient information delivery.
• Tool call executionIt supports multi-step tool calls and collaboration, laying the foundation for automating complex workflows.
• Database query generationIt accurately converts natural language into SQL query statements, enabling efficient retrieval of structured data.

Technical Principles of Yuan3.0 Ultra

• Layer-Adaptive Expert Pruning (LAEP)To address the issue of extremely uneven expert load during the pre-training of the MoE model, LAEP adaptively prunes low-load experts layer by layer during the training stabilization period and balances the load distribution among computing devices through a greedy rearrangement algorithm, reducing the model parameters from 1515B to 1010B, thus improving pre-training efficiency by 49%.
• Revised Reflection Inhibition Reward Mechanism (RIRM)In the rapid thinking reinforcement learning phase, the model is prone to overthinking. RIRM, through a reward constraint mechanism, gives higher rewards to correct samples with fewer reflection steps and heavier penalties to incorrect samples with more reflection steps, effectively suppressing the "overthinking" phenomenon. This results in a 16.33% improvement in training accuracy and a 14.38% reduction in output length.
• Localized Filtering-based Attention (LFA)It enhances semantic relationship modeling capabilities through local filtering mechanisms, achieving higher accuracy in capturing long-distance dependencies and fine-grained semantic associations compared to classic attention architectures.

Yuan3.0 Ultra project address

• GitHub repository: https://github.com/Yuan-lab-LLM/Yuan3.0-Ultra
• HuggingFace model libraryhttps://huggingface.co/YuanLabAI/Yuan3.0-Ultra

Application scenarios of Yuan3.0 Ultra

• Enterprise Knowledge Base Q&A: Build intelligent customer service and internal knowledge assistant based on private documents to achieve accurate retrieval and reliable answers.
• Intelligent analysis of financial documentsIt automatically parses complex documents such as financial statements, contracts, and approval forms, and completes data extraction, verification, and risk identification.
• Business Intelligence Report GenerationIt supports converting database query requirements into SQL statements and automatically generating visual data analysis reports.
• Smart office assistantIt supports routine office automation tasks such as automatic generation of meeting minutes, long document summaries, and email drafting.
• Multimodal Agent SystemAs the core engine driving complex intelligent agents, it coordinates tool calls, task planning, and multi-round interactions to automate business processes.