Introducing ChatGLM3, the Third-Generation Base Large Model by Zhipu AI, Achieving Multimodal Understanding and Code Generation

2023-10-30

At the 2023 China Computer Conference, Zhifu AI unveiled its self-developed third-generation base model, ChatGLM3, along with a series of related products. ChatGLM3 is a deeply optimized multimodal model capable of understanding and generating text, images, audio, and code. It has demonstrated excellent performance on publicly available datasets in both Chinese and English, with the MML, CEval, GSM8K, and BBH metrics improving by 36%, 33%, 179%, and 126% respectively. In addition to its multimodal understanding capabilities, ChatGLM3 can also generate and execute code, as well as enhance its knowledge through web searches. Furthermore, ChatGLM3 has improved its intelligent planning and execution abilities through its self-developed AgentTuning technology. ChatGLM3 has also introduced endpoint testing models for mobile and in-car platforms, providing efficient inference and cost-effective techniques. With the support of ChatGLM3 technology, the new generation of "Zhifu Qingyan" has become the first domestic large-scale model product capable of interacting with users through code, while also possessing search enhancement and image-text understanding capabilities. Zhifu AI has been committed to innovation and currently supports a variety of domestic hardware ecosystems. It is hoped that domestically developed large-scale models and chips can stand out on the international stage.