Zhipu releases the new generation flagship model GLM-5, focusing on enhancing programming and agent capabilities

On February 11, Zhipu officially launched the new flagship model GLM-5, focusing on programming and agent capabilities. The company claims to have achieved the best performance in the open-source domain. This marks another important release in the domestic AI large model Spring Festival lineup, following DeepSeek.

The parameter scale of GLM-5 has expanded from the previous generation’s 355B to 744B, with active parameters increasing from 32B to 40B. Zhipu confirmed that the mysterious model “Pony Alpha,” which topped the global model service platform OpenRouter’s popularity chart earlier, is actually GLM-5.

Internal evaluations show that GLM-5’s performance in front-end, back-end, and long-range programming development scenarios has improved by over 20% compared to the previous generation, with real programming experience approaching Claude Opus 4.5 level. The model is now available on chat.z.ai platform. This release signifies that domestic large models are continuously narrowing the gap with international leading standards in both technical approach and capability, providing developers with a new open-source option.

Parameter Scale Doubles, Pretraining Data Significantly Expanded

Zhipu’s new flagship model GLM-5 has achieved key upgrades at the architecture level. The parameter scale has expanded from 355B (activation 32B) to 744B (activation 40B), and pretraining data has increased from 23T to 28.5T, driven by larger-scale computing power investments that significantly enhance general intelligence capabilities.

The model introduces DeepSeek sparse attention mechanism for the first time, effectively reducing deployment costs and improving token utilization efficiency while maintaining long-text processing performance. This technical approach aligns with DeepSeek-V3/V3.2.

In terms of architecture, GLM-5 features 78 hidden layers, integrates 256 expert modules, activates 8 at a time, with approximately 44B active parameters, a sparsity of 5.9%, and a maximum context window of 202K tokens.

Significant Improvement in Programming Capabilities

The new flagship model GLM-5 demonstrates outstanding performance in internal Claude Code evaluations. In programming development scenarios such as front-end, back-end, and long-range tasks, it surpasses the previous GLM-4.7 comprehensively, with an average performance increase of over 20%.

GLM-5 can autonomously complete complex system engineering tasks such as agentic long-term planning and execution, backend refactoring, and deep debugging with minimal human intervention. The official states that the real programming experience has approached Claude Opus 4.5 level.

Zhipu positions GLM-5 as the latest flagship-level dialogue, programming, and agent model, emphasizing its enhanced capabilities in complex system engineering and long-range agent tasks.

Open-Source SOTA for Agent Capabilities

GLM-5 achieves open-source state-of-the-art (SOTA) in agent capabilities, ranking first among open-source models in multiple benchmarks. In BrowseComp (network retrieval and information understanding), MCP-Atlas (large-scale end-to-end tool invocation), and τ2-Bench (automatic agent planning and execution in complex scenarios), GLM-5 delivers the best performance.

To enable these capabilities, the model introduces a new “Slime” training framework supporting larger model architectures and more complex reinforcement learning tasks, significantly improving the efficiency of post-training reinforcement learning processes.

Additionally, Zhipu proposed an asynchronous agent reinforcement learning algorithm, enabling the model to continuously learn from long-term interactions and effectively unlock the deep potential of pretraining models. This mechanism has become one of the core technological features of GLM-5.

Domestic Large Model Releases During Spring Festival

Zhipu Qingyan GLM-5’s release marks the latest in a series of intense domestic AI large model launches during the Spring Festival period. On the same evening, Minimax also launched Minimax 2.5, just over a month after the previous version 2.2.

This wave of releases continues to heat up. Previously, DeepSeek introduced new models, and recent launches include Alibaba Qianwen’s Qwen 3.5, ByteDance’s SeeDance 2.0, among others. Multiple companies are simultaneously releasing new models during the Spring Festival window, reflecting that the competition in the domestic large model arena is entering a fierce stage.

Details of GLM-5 and Minimax 2.5’s technical documentation have not yet been fully disclosed, and their actual performance still awaits further validation by the developer community and professional institutions.

Risk Warning and Disclaimer

Market risks exist; investments should be cautious. This article does not constitute personal investment advice and does not consider individual users’ specific investment goals, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions herein are suitable for their particular circumstances. Invest at your own risk.