The fourth-generation CodeGeeX model, developed by ZhiPu AI, has been officially released and made open source.This model is the most powerful and versatile code generation large model with parameters below 10B, having been updated every six months since the first generation was released in 2014, and is dedicated to improving code generation capabilities.
CodeGeeX offers basic functions such as code completion, comments, fixes, and translation, as well as advanced features like connected Q&A, code interpreter, and enhanced search, all completely free for individual users and available in various mainstream IDEs. The number of individual users has exceeded 1 million, and the enterprise version is widely used in many industries and has received high ratings in the evaluation by the China Information Technology Institute.
The open-source version of the CodeGeeX4 series model, CodeGeeX4-ALL-9B, integrates various programming development functions and covers multiple programming scenarios. It has performed exceptionally well in several code evaluation sets and demonstrates strong code generation capabilities. The model supports 128K context, processes long code files, and achieves 100% accuracy in code retrieval tasks. Additionally, CodeGeeX4-ALL-9B is the only code large model to implement Function Call, with a success rate of over 90% in function call tests.
The CodeGeeX plugin version 2.12.0 has fully integrated the fourth-generation model, bringing faster code generation speed and various feature innovations, such as automatic generation of project README files, long text context memory, cross-file analysis and Q&A, local mode support, and enhanced NL2SQL capabilities. These features enhance the efficiency of code search, understanding, and maintenance, and improve the development experience.
Users can experience the latest technology by upgrading the CodeGeeX plugin in their IDE or downloading it from the plugin market. Meanwhile, ZhiPu AI has also released the enterprise-grade CodeGeeX4 generation all-in-one hardware and software product, providing code repository-level augmentation, Q&A, and enhanced search, supporting innovation and creation schemes, and suitable for more enterprise-level application scenarios.
Project Address: https://github.com/THUDM/CodeGeeX4