Zhipu officially released and open-sourced the new foundation model GLM-5 today. This model marks a major paradigm shift in large model capabilities: from simply writing code and front-end development to independently completing complex system engineering and long-term tasks through "Agentic Engineering." In the latest global authoritative Artificial Analysis ranking, GLM-5 ranks fourth globally and took the crown as the top open-source model.

QQ20260212-085506.png

Core Performance: Benchmarking Industry Leaders

GLM-5 demonstrates outstanding engineering capabilities in multiple benchmark tests, with its user experience approaching that of Claude Opus4.5:

  • Programming Alignment: Achieved SOTA performance in the open source community on SWE-bench-Verified (77.8) and Terminal Bench2.0 (56.2), surpassing Gemini3Pro.

  • Long-term Agent Capabilities: Maintained the top position among open-source models in BrowseComp, MCP-Atlas, and the complex tool planning test τ²-Bench.

  • Business Decision-Making: Demonstrated leading capabilities in long-term planning and resource management in the Vending Bench2 test simulating automated vending machine operations.

QQ20260212-085516.png

Technical Breakthroughs: Larger Scale and Higher Efficiency

The evolution of GLM-5 comes from a comprehensive upgrade of its underlying architecture:

  1. Parameter Expansion: The model expanded from 355B to 744B (40B activated parameters), with pre-training data increased to 28.5T.

  2. Asynchronous Reinforcement Learning: Adopted a new "Slime" framework and asynchronous agent reinforcement learning algorithm, allowing the model to continuously evolve from long-term interactions.

  3. Sparse Attention: Integrated the DeepSeek Sparse Attention mechanism, ensuring high-quality long-text processing while significantly optimizing Token efficiency and deployment costs.

Computing Power and Ecosystem: Deep Compatibility with Domestic Chips

In response to global computing power demands, GLM-5 has achieved deep inference compatibility with domestic computing platforms including Huawei Ascend, Moorten, Cambric, Kunlun, Muxi, Surpass, and HaiGuang. Through low-level operator optimization, it ensures high throughput and low latency operation on domestic chip clusters.

QQ20260212-085524.png

Application Scenarios: Reengineering the Development Process

GLM-5 is not just a model but also a "system architect." Zhipu simultaneously launched the Z Code programming tool based on it, supporting multiple agents to complete the entire closed-loop process from task decomposition, code writing, to debugging preview. In addition, GLM-5 also supports directly converting materials into office documents such as .docx, .pdf, and .xlsx, and provides native AI plug-in support in the Excel environment.

Currently, the model weights of GLM-5 have been open-sourced through Hugging Face and ModelScope, and are subject to the MIT License.