Z.AI: GLM 4.7
Model Type
Proprietary Model
API access only
Recommended Use Cases
Text Generation
Try GLM 4.7
GLM-4.7 is Z.AI's coding-focused flagship model, designed for real-world development workflows with strong multi-step task execution, tool use, and Claude Code compatibility.
GLM-4.7 achieves 87.4% on τ²-Bench—the highest reported result among publicly available open-source models to date. — Z.AI
Overview
Released December 22, 2025, GLM-4.7 builds on GLM-4.6 with significant improvements for production coding environments. Unlike chat-focused predecessors, it targets lengthy task cycles, frequent tool use, and multi-step consistency. Selected as the default model for Z.AI's GLM Coding Plan.
Key Capabilities
- 200K context window with 128K output tokens
- Preserved Thinking: Retains reasoning across turns, avoiding context collapse
- Turn-level Thinking Control: Enable/disable reasoning per turn for speed vs. accuracy
- Vibe Coding: Generates cleaner, modern UIs with improved aesthetics
- Multi-framework support: Claude Code, Cline, Roo Code, Kilo Code, TRAE
Performance Highlights
| Benchmark | GLM-4.7 | vs GLM-4.6 |
|---|---|---|
| SWE-bench Verified | 73.8% | +5.8% |
| SWE-bench Multilingual | 66.7% | +12.9% |
| Terminal Bench 2.0 | 41.0% | +16.5% |
| τ²-Bench | 87.4% | Best open-source |
| BrowseComp | 67.5% | Strong web tasks |
| HLE | 42.8% | +12.4% |
When to Use GLM-4.7
Choose GLM-4.7 when you need:
- Production coding workflows with multi-step stability
- Compatibility with popular coding agents (Claude Code, Cline, etc.)
- Multilingual software engineering beyond Python
- Terminal-based agentic tasks
- "Vibe coding" with polished UI generation
Choose GLM-4.7 Flash when you need:
- Local deployment on consumer GPUs (RTX 3090/4090)
- Lower latency for simpler coding tasks
- Free-tier API access
Choose GLM-5 when you need:
- Maximum capability for complex systems engineering
- Long-horizon agentic tasks requiring deeper reasoning
- Lower hallucination rates
Role in Series
GLM coding model evolution:
- GLM-4.5 (Jul 2025): First agent-native model, 64.2% SWE-bench
- GLM-4.6 (Sep 2025): 200K context, 68% SWE-bench
- GLM-4.7 (Dec 2025): Preserved Thinking, 73.8% SWE-bench (this model)
- GLM-5 (Feb 2026): 77.8% SWE-bench, flagship