What is the Congrong Large Model?
The Congrong Large Model is a multimodal AI model launched by CloudWalk Technology. On the internationally authoritative evaluation platformOpenCompass, the model topped the multimodal leaderboard with a score of 80.7, surpassing top teams such as Google and OpenAI. The model focuses on general visual-language understanding and reasoning tasks. Based on core technological breakthroughs such as multimodal alignment, human-like decision-making, efficient engineering optimization, and native multimodal reasoning, it builds a globally leading technological barrier. The Congrong Large Model performs outstandingly in many fields such as medical health, mathematical logic, and art design, and has achieved large-scale implementation in multiple fields such as finance, manufacturing, and government affairs, helping intelligent transformation.
Main functions of the Congrong Large Model
- Visual perception and cognitive understanding: Supports processing visual information (such as images and videos) for cognitive understanding, for example, performing outstandingly in fields such as medical health and art design, and being able to understand complex visual scenes.
- Cross-domain applications: Demonstrates strong understanding and reasoning capabilities in multiple professional fields (such as mathematical logic, medical health, and art design).
- Text recognition in complex scenes: Performs text recognition in complex scenes (such as OCRbench), supports processing high-resolution images and documents (such as contracts and invoice tables), and supports tasks such as intelligent review, intelligent parsing, and intelligent question answering.
- Open-domain question answering: Performs excellently in open-domain question answering (such as MMVet), providing accurate and in-depth answers.
Technical principles of the Congrong Large Model
- Multimodal alignment: Builds high-quality benchmark datasets covering multiple task scenarios, and improves the model's understanding and reasoning capabilities for multimodal data based on reinforcement instruction alignment. It integrates DPO and GRPO technologies to optimize the model's learning mechanism, making the model's decision-making and reasoning closer to human thinking, and achieving human-like reasoning and decision-making without relying on a reward model.
- Efficient engineering optimization: For high-resolution images and multimodal document understanding tasks, it optimizes the structure of the model's image encoder to efficiently process high-resolution images and complex documents. It optimizes the model's context modeling capability to accurately track logical relationships in long texts, supporting tasks such as cross-page document analysis and multi-turn dialogue.
- Native multimodal reasoning: Upgrades the model architecture to handle image-text interleaved patterns in multi-image and cross-image scenarios and native video modes, enabling complex multimodal tasks such as cross-image comparison, image-text combined reasoning, and multi-image question answering.
Application scenarios of the Congrong Large Model
- Financial risk control: Cooperates with banks to build risk control AI agents, automate risk identification, and reduce complaint volume.
- Intelligent customer service: Deploys an intelligent customer service platform for e-commerce platforms, improving question-answering accuracy and customer service efficiency.
- Medical health: Processes medical images, assists doctors in diagnosis, and improves diagnostic accuracy and efficiency.
- Government affairs: Processes government documents, enables intelligent review and question answering, and optimizes public services.
- Manufacturing: Used in product quality inspection to improve production efficiency and product quality.