Zhipu AI Launches Domestic Large Model GLM-4 to Rival OpenAI
Credit: CFP
BEIJING, January 18 (TMTPOST)—Zhipu AI, an AI star startup backed by Alibaba and Tencent, has introduced its latest large model GLM-4 with an aim to compete with OpenAI’s GPT-4.
The product was unveiled during the "Zhipu DevDay" event in Beijing on Tuesday.
The GLM-4 model, whose performance the company claims parallels GPT-4, features a longer context window, improved multimodal capabilities, and faster inference speed. It is designed for higher concurrency, aiming to reduce inference costs, making it a potentially more cost-effective option for businesses.
A notable enhancement in GLM-4 is its improved intelligent agent capabilities. The model introduces GLM-4 All Tools, a feature that autonomously comprehends and plans complex instructions based on user intent. This enables it to interact with web browsers, interpret code, and collaborate with multimodal text-generation large models to accomplish intricate tasks.
GLM-4 can handle a context window length of 128k and process a text equivalent to 300 pages with a single prompt. In the needle test, a measure of information retrieval accuracy, GLM-4 reportedly has achieved almost 100% accuracy within a 128K text length. The model also demonstrates enhanced multimodal capabilities, including text generation and multimodal understanding.
The All Tools capability of GLM-4, a novel feature, harnesses the model's robust agent abilities. It can autonomously interpret and plan complex instructions, interacting with web browsers, code interpreters, and the CogView3 model of text generation. Furthermore, it can automatically call a code interpreter for solving complex equations or calculus problems. When tested on the GSM8K, Math, and Math23K datasets, GLM-4 All Tools achieved results on par with GPT-4 All Tools.
In addition to its capabilities, the All Tools feature of GLM-4 is fully automatic and versatile, capable of handling a wide range of tasks such as file processing, data analysis, and chart drawing. It can also handle various file formats, including Excel, PDF, and PPT.
The company has raised over 2.5 billion yuan (US$ 350 million) in funding in 2023, and its current valuation has exceeded tens of billions of yuan, making it a unicorn in China’s AI field. It has received investments from prominent institutions such as the Social Security Fund Zhongguancun Independent Innovation Fund, Meituan, Alibaba, Tencent, Xiaomi, Kingsoft, Sequoia, Gaorong, and others.
Zhipu AI has also collaborated with dozens of companies, including Deloitte, Mengniu, China Telecom, Kingsoft, Mango TV and Huatai Securities. Their cooperation spans various industries, including government affairs, finance, energy, manufacturing, and healthcare, supporting over 10 domestic hardware ecosystems. Zhipu AI’s CEO Zhang Peng revealed that in the second half of 2023, the company secured contracts with hundreds of clients, with an overall amount under contracts reaching the billion-yuan level.
Zhang emphasized the company’s commitment to the long-term development and innovation of large models. He expressed hope for collaboration with all stakeholders in the AI industry, including partners in the industrial chain, developer communities, and academia, to contribute to China’s AI future.
Addressing the gap between Chinese large models and GPT, Zhang acknowledged that they are still in a catching-up phase, mainly due to the late start of domestic large models compared to OpenAI. He emphasized the need for continuous accumulation and adjustment of development speed and trajectory through innovations in algorithms, system engineering, data, and application landing to surpass GPT.