GLM-4.7

by Z.ai (GLM project)

Next-generation open-source coding model with advanced thinking capabilities and MoE architecture

See vendor docs and community benchmarks

Summary

GLM-4.7 is Z.ai’s latest flagship model released December 2025, built specifically for real development workflows. It features enhanced programming capabilities, more stable multi-step reasoning, and introduces three levels of thinking: Interleaved, Preserved, and Turn-level. The model uses a Mixture-of-Experts (MoE) architecture with 200k token context window and 128k token output capacity.

Features

  • 200k token context window with 128k max output tokens
  • Mixture-of-Experts (MoE) architecture for computational efficiency
  • Interleaved Thinking: Automatic reasoning before responses and tool calls
  • Preserved Thinking: Retains reasoning across multi-turn conversations in coding agents
  • Turn-level Thinking: Per-turn control over reasoning (enable for complex tasks, disable for simple requests)
  • Enhanced multilingual coding support with strong terminal-based task performance
  • Advanced “Vibe Coding” for UI/webpage generation with cleaner, modern outputs
  • Superior tool use capabilities (87.4 on τ²-Bench, highest among open-source models)

Superpowers

GLM-4.7 excels at agentic coding workflows with significant improvements in SWE-bench (73.8%, +5.8% over GLM-4.6), SWE-bench Multilingual (66.7%, +12.9%), and Terminal Bench 2.0 (41%, +16.5%). Integrates seamlessly with Claude Code, Kilo Code, Cline, and Roo Code frameworks. Ranks #1 among open-source models on Code Arena and #6 overall in WebDev benchmarks. Strong mathematical reasoning with 42.8% (+12.4%) on HLE benchmark.

Pricing & access

  • Available via Z.ai API Platform
  • Open-source weights on HuggingFace and ModelScope
  • Supports vLLM and SGLang inference frameworks for local deployment
  • Accessible through OpenRouter

Known limitations & notes

  • Requires substantial infrastructure for self-hosting (GPU RAM, NVMe for 200k context)
  • Turn-level thinking requires manual configuration for optimal latency vs accuracy tradeoff
  • Best performance achieved when thinking is enabled for complex tasks

Sources / notes: