Zhipu AI CEO Zhang Peng: Building a Groundbreaking Large Model, the Fully Self-Developed GLM-4 Launches Today
-
On January 16, Zhipu AI's first Technology Open Day (Zhipu DevDay) was successfully held at the Zhongguancun National Innovation Demonstration Zone Exhibition Center in Beijing. At the event, the Zhipu AI team comprehensively showcased the technological achievements accumulated over more than three years of dedication to large model development and officially released the new-generation foundational large model, GLM-4.
The launch of GLM-4 marks a critical breakthrough in Zhipu AI's large model research and development. GLM-4's overall performance has significantly improved compared to its predecessor, approaching the level of GPT-4. It supports longer context windows, features enhanced multimodal capabilities, and delivers faster reasoning speeds with higher concurrency, substantially reducing inference costs. Additionally, GLM-4 greatly improves agent capabilities. GLM-4 All Tools can autonomously understand and plan complex instructions based on user intent, freely invoking tools such as a web browser, Code Interpreter, and multimodal text-to-image models to accomplish complex tasks. The GLMs personalized agent customization feature has also been launched, allowing users to create their own GLM agents with simple prompt instructions.
Specifically, GLM-4 supports a context window length of 128k, enabling it to process up to 300 pages of text in a single prompt. In the needle test, the GLM-4 model achieved nearly 100% precision recall within the 128K text length, with no decline in accuracy due to loss of focus in long-context global information. Moreover, its multimodal capabilities, including text-to-image generation and multimodal understanding, have been significantly enhanced. Additionally, GLM-4's All Tools capability has been officially released. Leveraging the powerful Agent capabilities of the GLM model, GLM-4 can autonomously understand and plan complex instructions based on user intent, freely utilizing web browsers, the Code Interpreter, and the text-to-image CogView3 model. Moreover, GLM-4 automatically employs the code interpreter to solve complex equations or calculus problems. When compared across the GSM8K, Math, and Math23K datasets, GLM-4 All Tools achieves results comparable to GPT-4 All Tools. The All Tools capability is fully automated and can handle various tasks, including file processing, data analysis, chart creation, and supports file formats such as Excel, PDF, and PPT.
GLM's personalized agent customization feature has also been launched. Based on the robust capabilities of the GLM-4 model, users can create their own GLM agents simply by logging into the official ChatGLM website (https://www.chatglm.cn/) and using simple prompt instructions. Additionally, users can share their created agents through the newly launched Agent Center. The introduction of GLM model agents signifies that anyone can freely utilize the GLM-4 model and explore its potential, enabling convenient development of large models without any programming background. This represents Zhipu AI's latest effort to lower the barrier to using large models.
Previously, Zhipu AI collaborated with the CCF (China Computer Federation) and the Chinese Information Processing Society's Social Media Processing Committee to establish the CCF-Zhipu Large Model Fund and the SMP-Zhipu Large Model Interdisciplinary Fund. These funds support research and cross-disciplinary innovation related to large model theory, algorithms, models, and applications. Last year, the two funds provided over 10 million RMB in cash and computing resources to support 41 research projects involving more than 30 universities across China. In 2024, Zhipu AI will launch an open-source fund for large models, which includes three "thousand" initiatives: Zhipu AI will provide the open-source community with a thousand GPUs to facilitate open-source development; allocate 10 million RMB in cash to support open-source projects related to large models; and offer 100 billion free API tokens to outstanding open-source developers. Zhang Peng stated that the purpose of the large model open-source fund is to drive significant progress in large model research and development and to promote the prosperity of the entire open-source ecosystem for large models.
Facing global large model entrepreneurs, Zhipu AI will also upgrade its "Z Plan," collaborating with ecosystem partners to establish a 1 billion RMB large model startup fund. This fund will support original innovations in large models, covering areas such as large model algorithms, underlying operators, chip optimization, industry-specific large models, and super applications.
2024 is widely regarded as another pivotal year in the history of artificial intelligence development. At this juncture, by hosting a technology open day, Zhipu AI not only showcases its leading position in the AI field but also injects new momentum into the industry's growth. Zhang Peng, CEO of Zhipu AI, said, "The innovative research and development of large models is a long march. A single flower does not make spring. We hope to work together with all participants in the AI industry, partners across the supply chain, developer communities, and academia to contribute our modest efforts to the new future of AI in China."