Hugging Face Overview
GLM-5: Advanced Agentic AI for Complex Systems
GLM-5 is a significant advancement in large language models, developed by zai-org, specifically engineered for complex systems engineering and long-horizon agentic tasks. It scales to 744 billion parameters (with 40 billion active parameters) and has been pre-trained on an extensive 28.5 trillion tokens of data.
Key Innovations & Capabilities
- DeepSeek Sparse Attention (DSA): Integrates DSA to substantially reduce deployment costs while preserving its long-context processing capabilities.
- Asynchronous RL Infrastructure (slime): Utilizes a novel
slimeinfrastructure for asynchronous reinforcement learning, significantly improving training throughput and efficiency for fine-grained post-training iterations. - Best-in-Class Open-Source Performance: Achieves leading performance across a wide range of academic benchmarks, particularly in reasoning, coding, and agentic tasks, closing the gap with frontier models.
Benchmark Highlights
GLM-5 demonstrates strong results across various challenging benchmarks, including:
- HLE (Humanity's Last Exam): Scores 30.5 (text-only) and 50.4 (with tools), indicating advanced reasoning capabilities.
- SWE-bench Verified: Achieves 77.8% on verified software engineering benchmarks.
- Terminal-Bench 2.0: Scores 56.2 / 60.7 on Terminus 2 and 56.2 / 61.1 on Claude Code, showcasing robust performance in terminal-based tasks.
- BrowseComp: Reaches 62.0% (without context management) and 75.9% (with context management), highlighting its ability to navigate and interact with web environments.
Deployment
GLM-5 supports local deployment via popular frameworks like vLLM, SGLang, KTransformers, and xLLM, offering flexibility for developers.