Groundbreaking advancements continue to emerge in the field of artificial intelligence and large language models. Z.ai's latest next-generation large language model, GLM-4.7, is designed to cater to domains requiring more complex and long-term task cycles, such as software development, compared to earlier models primarily focused on single-turn interactions. This new model has been engineered to meet critical developer needs, including offline operation, frequent tool usage, and uninterrupted stability.
The capabilities of GLM-4.7 have been proven through various comprehensive tests. Particularly, in evaluations conducted on 100 real coding tasks within a Claude Code development environment, the model demonstrated impressive performance in areas like frontend, backend, and instruction following. Achieving a score of 87.4 on platforms like τ²-Bench, which measure interactive tool usage, GLM-4.7 secured the highest score among publicly available open-source models to date. This achievement is recognized as a testament to the model's ability to understand and manage complex workflows.
Even more noteworthy is GLM-4.7's performance on Code Arena, a large-scale blind evaluation platform. With over a million participants, GLM-4.7 not only secured the leadership position among open-source models but also ranked highest among models developed in China. These results underscore Z.ai's commitment to its open-source philosophy by developing systems that can be reliably used in real projects. GLM-4.7's ability to outperform strong competitors like GPT-5, Gemini, and Claude in numerous academic and practical tests is creating a new competitive dynamic in the AI ecosystem.
Among the innovations brought by GLM-4.7, significant improvements are evident in multilingual coding and terminal-based tasks compared to its predecessor, GLM-4.6. With a success rate of 73.8% on SWE-bench and reaching 66.7% on SWE-bench Multilingual, the model demonstrates cleaner coding capabilities that overcome language barriers. Furthermore, considerable progress has been made in complex tasks thanks to its "thinking before acting" capability. These advancements have the potential to transform GLM-4.7 from a mere tool into an indispensable partner for developers.