Skip to content

Latest commit

 

History

History
34 lines (24 loc) · 1.17 KB

File metadata and controls

34 lines (24 loc) · 1.17 KB

Overview

GLM-4.7 is a frontier-level open-source model from Zhipu AI featuring a Mixture-of-Experts (MoE) architecture designed to balance computational efficiency with depth. The model represents cutting-edge performance in code generation and reasoning tasks.

Key Specifications

  • Context Window: 200,000 tokens
  • Output Capacity: 128,000 tokens per generation
  • Architecture: MoE-based for efficiency
  • Can generate complete software frameworks in a single pass

Benchmark Performance

Coding Excellence:

  • SWE-bench: 73.8% (5.8% improvement over predecessor)
  • SWE-bench Multilingual: 66.7% (12.9% improvement)
  • LiveCodeBench-v6: 84.9% (surpassing Claude Sonnet 4.5 at 64%)
  • Terminal Bench 2.0: 41% (+16.5% improvement)

Reasoning & Tool Use:

  • τ²-Bench: 87.4 (highest for any open-source tool-using model)
  • HLE Benchmark: 42.8% (+12.4% improvement)
  • AIME 2025: 95.7%
  • GPQA-Diamond: 85.7%

Advanced Features

  • Interleaved Thinking with Preserved Thinking and Turn-level Thinking
  • Tool-use capabilities for complex workflows
  • Local deployment support via vLLM and SGLang

Deployment

Supports multiple inference frameworks for flexible local and cloud deployment.