GLM-4.7

GLM-4.7 adopts a Mixture-of-Experts (MoE) architecture that activates only specialized subnetworks for each task, keeping computational overhead manageable while preserving deep reasoning capability across diverse problems. The model was built as a coding partner with a particular emphasis on agentic development workflows—multilingual code generation, terminal-based task execution, and UI quality improvement are core strengths. Its 200,000-token context window and massive output capacity allow developers to generate entire software frameworks or lengthy technical documents in a single pass. The model introduces capabilities the developers describe as Interleaved Thinking, Preserved Thinking, and Turn-level Thinking—design choices meant to keep behavior stable and controllable across long, multi-step software tasks.

The model builds on its predecessor GLM-4.6 with targeted improvements on mainstream agent frameworks including Claude Code, Kilo Code, Cline, and Roo Code. Benchmark results show meaningful gains: 73.8% on SWE-bench (+5.8%), 66.7% on the multilingual variant (+12.9%), and a 16.5-point jump on Terminal Bench 2.0, indicating real-world coding utility rather than synthetic test performance. Complex reasoning also advances significantly—GLM-4.7 reaches 42.8% on the HLE (Humanity's Last Exam) benchmark, a 12.4-point improvement over its predecessor, suggesting stronger mathematical and logical problem-solving. Tool use capabilities are similarly sharpened, with measurable improvements on τ²-Bench and web browsing tasks. Open-sourced and commercially available, the model targets developers who want frontier-level coding and reasoning without lock-in to closed APIs.