GLM-4.7

GLM-4.7 is a flagship open-source coding model from Z.ai, a company spun out of Tsinghua University, designed from the ground up for autonomous software development rather than casual chat interactions. With 355 billion total parameters and 32 billion active parameters during inference, the model balances scale with practical efficiency. What sets it apart is its architecture for agentic workflows—specifically engineered for terminal-based coding agents that need to maintain context across long, multi-file programming sessions. Unlike models that restart reasoning from scratch each turn, GLM-4.7 preserves its thinking blocks across conversations, enabling genuine continuity in complex coding tasks that span multiple files and sessions.

The model has quickly proven itself on competitive benchmarks, achieving 73.8% on SWE-bench Verified, 87.4% on τ²-Bench, and 84.9% on LiveCodeBench—placing it among the top open-source models on LMArena Code Arena and competitive with proprietary alternatives at a fraction of the cost. It incorporates advanced interleaved thinking for stable multi-step reasoning, making it well-suited for production agentic deployments. Originally released in December 2025, GLM-4.7 ships under a MIT license, giving developers full freedom to integrate it into coding agents like Claude Code, Cline, and Roo Code. The combination of its benchmark performance, multi-turn stability, and open-source accessibility has made it a practical choice for developers seeking a capable alternative to closed models for automated coding workflows.