GLM-4.7 (Reasoning) is a large-scale Mixture-of-Experts (MoE) language model developed by Zhipu AI (Z.ai), specifically optimized for complex reasoning, mathematics, and agentic coding tasks. Building upon the GLM-4 series, this model incorporates reinforcement learning-driven reasoning paradigms, including Interleaved Thinking and Preserved Thinking, which allow it to plan, verify, and maintain logical consistency across multi-step execution processes.
The model is characterized by its high performance in STEM-related benchmarks and its ability to handle long-horizon tasks. It features a context window of 200,000 tokens, facilitating the processing of extensive code repositories and large-scale technical documentation. Its architecture is designed to rival proprietary frontier models while maintaining an open-weight distribution for research and deployment.
Technical Features
- Thinking Modes: Introduces specialized modes that enable the model to "think" before tool invocation or generating responses, improving instruction following and error recovery.
- MoE Architecture: Utilizes a 358B parameter Mixture-of-Experts structure, balancing computational requirements with the depth of knowledge necessary for specialized professional domains.
- Coding Proficiency: Features enhanced capabilities in multilingual agentic coding, achieving competitive scores on benchmarks such as SWE-bench and LiveCodeBench.