Zhipu AI presented its cutting-edge large model, GLM-4, at the inaugural Technology Open Day in Beijing. Developed by a team from Tsinghua University, GLM-4 showcases significant advancements, particularly in Chinese language capabilities, approaching the level Chat GPT-4
The event highlighted GLM-4’s capabilities, including support for longer context, faster inference speed, and improved intelligence, allowing autonomous understanding and planning of complex instructions based on user intent.
Notably, GLM-4 can handle a context window length of 128k and achieve almost 100% accuracy recall in the Needle in the Haystack test within a text length of 128k. This test, designed by AI entrepreneur Greg Kamradt, is considered a benchmark for large model accuracy in long texts.
Zhipu AI emphasized GLM-4’s role in reducing inference costs, with its applications ranging from accessing web browsers to code interpreters and multimodal models.
The presentation unveiled GLM-4’s personalized intelligent agent customization feature, allowing individuals without coding skills to develop customized AI agents.
These agents can serve diverse roles, such as optimizing emails for white-collar workers, designing engaging teaching plans for teachers, or quickly creating logos based on keywords.
CEO Zhang Peng underscored the company’s commitment to lowering the threshold for large model development, emphasizing that anyone can explore GLM-4’s potential without a programming background.
The global impact of Zhipu AI’s large models was highlighted, with over 10 million downloads of the open-source versions, making them the most downloaded and influential domestic large models.
Beijing stands out as a hub for large model innovation, boasting over 120 entities engaged in this field, with a significant majority being enterprises.
The Haidian district alone hosts more than 80 of these entities, representing around 70% of the city’s total. This underscores Beijing’s leading position in large model research and development within the country.
