Zhipu releases the new generation flagship model GLM-5, with a focus on improving programming and agent capabilities.

```

On February 11th, Zhipu officially launched its next-generation flagship model, GLM-5, focusing on programming and agent capabilities. The company claims it has achieved the best performance in the open-source sector. This marks another major release of Chinese AI large models during the Lunar New Year, following DeepSeek.

GLM-5 expands its parameter scale from 355B in the previous generation to 744B, and activation parameters from 32B to 40B. Zhipu confirmed that the mysterious model "Pony Alpha," which topped the global model service platform OpenRouter's popularity ranking, is GLM-5.

Internal evaluations show that GLM-5 delivers an average performance improvement of over 20% compared to the previous generation in programming and development scenarios such as front-end, back-end, and long-range tasks. Its real coding experience is close to Claude Opus 4.5. The model is now available on the chat.z.ai platform. This release marks the continued narrowing of the gap between domestic large models and leading international standards in technical routes and capabilities, providing developers with a new open-source option.

Parameter Scale Doubled, Pre-training Data Significantly Expanded

Zhipu's new flagship model GLM-5 achieves key upgrades at the architectural level. Parameter scale is expanded from 355B (activation 32B) to 744B (activation 40B), with pre-training data increased from 23T to 28.5T. Larger computational power input significantly enhances general intelligence capabilities.

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

In terms of architecture configuration, GLM-5 builds 78 hidden layers, integrates 256 expert modules (activating 8 at a time), with about 44B activation parameters and 5.9% sparsity, supporting up to 202K tokens in the context window.

Significant Improvement in Programming Capabilities

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

GLM-5 can autonomously complete complex system engineering tasks such as agentic long-range planning and execution, back-end reconstruction, and deep debugging with minimal human intervention. Officially, the real programming experience is already close to Claude Opus 4.5.

Zhipu positions GLM-5 as a flagship conversation, programming, and agent model, focusing on strengthening its capabilities in complex system engineering and long-range agent tasks.

Agent Capabilities Achieve Best Open-Source Performance

GLM-5 achieves SOTA (state-of-the-art) in agent capabilities in open-source, ranking first in multiple evaluation benchmarks. In three tests—BrowseComp (networked retrieval and information understanding), MCP-Atlas (large-scale end-to-end tool calling), and τ2-Bench (automatic tool planning and execution in complex scenarios)—GLM-5 achieved the best performance.

To achieve breakthroughs in capabilities, the model constructed a new "Slime" training framework, supporting larger scale model architectures and more complex reinforcement learning tasks, significantly improving the efficiency of post-reinforcement training processes.

Additionally, Zhipu introduced an asynchronous agent reinforcement learning algorithm, enabling the model to continuously learn from long-range interactions and effectively unlock the deep potential of the pre-trained model. This mechanism has become one of GLM-5’s core technical features.

Intensive Releases of Domestic Large Models During Lunar New Year

The release of Zhipu Qingyan GLM-5 has become the latest highlight among the intense competition of Chinese AI large models during the Lunar New Year. 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. DeepSeek recently launched a new model, Alibaba's Qwen 3.5, and ByteDance's SeeDance 2.0 have also been unveiled. Multiple vendors have chosen to launch new products during the New Year window, reflecting that the competition in domestic large model tracks is entering a white-hot stage.

Detailed technical documentation for GLM-5 and Minimax 2.5 has not yet been fully disclosed, and their actual performance awaits further verification by the developer community and professional institutions.

Risk Warning and DisclaimerThe market has risks; investment should be cautious. This article does not constitute personal investment advice and does not take into account individual users' special investment objectives, financial status, or needs. Users should consider whether the opinions, viewpoints, or conclusions in this article are suitable for their specific circumstances. Investments based on this information are their own responsibility. ```