GLM-4.7

GLM-4.7 builds on the GLM-4.x family foundation using a Mixture-of-Experts Transformer architecture that routes tasks through specialized expert pathways, keeping the core design from the previous version while updating the weights to improve performance. This architecture enables efficient handling of complex, multi-step tasks typical in real development environments, where developers need a model that remains consistent across lengthy task cycles, calls the right tools reliably, and produces stable behavior when interacting with external systems. The model ranks as the top open-source option on the Artificial Analysis Intelligence Index and leads on benchmarks like tau-bench and SWE-bench, making it particularly strong for production coding workflows and agentic execution.

Released under a permissive MIT-style license, GLM-4.7 is designed to be fine-tuned, self-hosted, and flexibly deployed across different infrastructure. The model shows significant improvements over GLM-4.6 in coding capabilities and multi-step reasoning, while also delivering more natural conversational output for writing and role-playing scenarios. It supports up to 200K context tokens with an output capacity reaching 128K tokens, making it suitable for complex projects that require extended reasoning chains. The combination of benchmark leadership, open-weight availability, and tool-use stability positions GLM-4.7 as a practical choice for developers building AI-native applications, coding agents, and automated development pipelines.