Introduction

This project established a new-generation technology system for cognitive large model, developed the GLM series of models, and the main performance indicators are aligned with the international advanced levels. The series of models are open-source and achieved large-scale applications. The project has generated extensive international influence and achieved significant scientific, economic, and social benefits.

Established a Cognitive Large Model Technology System, with Key Performance Indicators Aligned with International Advanced Levels

To address challenges such as large model framework design, content understanding and generation, and complex planning, this project has established a new-generation technology system for cognitive large model, and the main performance indicators are aligned with the international advanced levels.

In terms of model architecture, a general pre-training architecture GLM based on autoregressive blank infilling was proposed. A masked attention mechanism is integrated into the autoregressive generation process to enhance the robustness and reduce hallucination of large model.

In terms of multimodal models, a multimodal understanding and generation algorithm based on visual experts is proposed, which deeply aligns language and multimodal signals to achieve multimodal cognition and interaction in large models.

In terms of agent models, an autonomous tool-using agent technology based on deep reasoning has been developed, enabling the intelligent planning and execution of tasks across general graphical interfaces, mobile phones, and computers.

Established GLM Series of Models and Diverse Service Patterns, Achieved Large-Scale Applications

This project developed a series of GLM models, including foundation models, dialogue models, multimodal models, and agent models. These models provide services for multimodal content understanding and generation (encompassing text, image, video, etc.), as well as autonomous execution services by agents. The project has developed the dialogue model ChatGLM and built the generative AI service assistant “Zhipu Qingyan”. A one-stop large model development platform based on Model as a Service (MaaS) has been established, which provides services through API calls, cloud-based fine-tuning, cloud-based private deployment, and other methods. The daily average API calls exceed 1 trillion characters. The key technologies of the project have been widely applied. A series of products and services have achieved large-scale applications in over 10,000 organizations across more than 20 industries, including banking, telecommunications, government services, and e-commerce. This has helped industries develop successful large model applications, promoted their intelligent upgrading. The project has achieved significant scientific, economic, and social benefits.

Practicing the Philosophy of Large Model Open-Source to Foster Technological Development, Gained High International Recognition

The project emphasizes the development of the large model ecosystem. It has open-sourced over 40 models, which have garnered more than 150,000 stars on GitHub. They have spawned over 1,000 open-source projects, fostering technological development in various fields. We took the lead to open-source the ChatGLM-6B series of models, which ranked first on the global large model trend list of Hugging Face for 4 weeks. Additionally, our team ranked 5th among the "The top 15 most-liked organizations on Hugging Face" in 2023, it is the only Chinese organization on the list. The international top journal Nature introduced the practices of the ChatGLM model in developing Chinese large language models in the article "China's ChatGPT: why China is building its own AI chatbots”, which has attracted widespread attention globally. Furthermore, our team delivered invited reports at top international AI conferences, including ICLR 2024 and WWW 2024.