GLM-4.7

Core Coding Capabilities:
• #1 open-source model on LMArena Code Arena (outperforming GPT-5.2)
• 73.8% on SWE-bench Verified (+5.8% over GLM-4.6)
• 66.7% on SWE-bench Multilingual (+12.9% improvement)
• 84.9% on LiveCodeBench-v6 (surpassing Claude Sonnet 4.5 at 64%)
• 41% on Terminal Bench 2.0 (+16.5% improvement)

Advanced Reasoning:
• AIME 2025: 95.7% (open-source SOTA)
• GPQA-Diamond: 85.7%
• HLE with Tools: 42.8% (+12.4% over GLM-4.6)
• HMMT Feb 2025: 97.1%
• MMLU-Pro: 84.3%

Agent & Tool Capabilities:
• τ²-Bench: 87.4% (approaching Claude Sonnet 4.5 at 87.2%)
• BrowseComp: 52% (67.5% with context management)
• Significantly improved tool-calling and web browsing performance

Interleaved Thinking System:
• Interleaved Thinking: Thinks before every response and tool call for improved instruction following
• Preserved Thinking: Automatically retains thinking blocks across multi-turn conversations, reducing information loss
• Turn-level Thinking: Per-turn control over reasoning—disable for lightweight requests, enable for complex tasks
• Optimal for long-horizon, complex agentic workflows

Vibe Coding:
• Enhanced UI quality with cleaner, more modern webpages
• Better-looking slides with accurate layout and sizing
• Improved frontend aesthetic generation for low-code platforms and rapid prototyping

Architecture:
• 200K context window with 128K maximum output tokens
• MIT licensed, open weights
• Supports vLLM, SGLang, and Transformers inference frameworks

‍

Agentic Coding & Development:
• End-to-end software engineering from requirement comprehension to executable code
• Multilingual agentic coding across multiple programming languages
• Terminal-based development tasks with stable multi-step execution

Frontend & UI Generation:
• Production-quality web UI generation with enhanced aesthetics
• Modern webpage creation with clean layouts and accurate sizing
• Slide and poster generation with professional design consistency
• Low-code platforms and AI frontend generation tools

Complex Agent Workflows:
• Long-horizon tasks requiring preserved reasoning across turns
• Web browsing and information retrieval with context management
• Tool-calling and function execution for enterprise automation
• Real-world interaction scenarios (τ²-Bench: retail, telecom, airline)

Enterprise Applications:
• Development support and solution discussions with context-aware responses
• Decision-making assistance with advanced reasoning capabilities
• Mathematical problem-solving and scientific reasoning
• Creative writing, role-play, and conversational AI with improved quality

Research & Prototyping:
• Rapid prototyping with superior UI aesthetics and accurate implementation
• Complex demos and proof-of-concept development
• Educational tools and interactive learning applications with 200K context support

‍