KNOWLEDGE ATLAS released the new generation flagship model GLM-5, focusing on enhancing programming and intelligent agent capabilities

On February 11, KNOWLEDGE ATLAS officially launched the new generation flagship model GLM-5, focusing on programming and agent capabilities, with the official claim of achieving optimal performance in the open-source field. This is another significant release of a domestic AI large model during the Spring Festival, following DeepSeek. The parameter scale of GLM-5 has expanded from the previous generation's 355B to 744B, with the activation parameters increased from 32B to 40B. KNOWLEDGE ATLAS confirmed that the mysterious model "Pony Alpha," which previously topped the popularity chart on the global model service platform OpenRouter, is indeed GLM-5. Internal evaluations show that GLM-5 has improved average performance by over 20% in programming development scenarios such as front-end, back-end, and long-range tasks compared to the previous generation. The real programming experience approaches the level of Claude Opus 4.5. This model has been launched on the chat.z.ai platform. This release marks a continuous narrowing of the gap between domestic large models and international leading levels in terms of technical paths and capability performance, providing developers with new open-source options. Parameter Scale Doubled, Pre-training Data Significantly Expanded The new generation flagship model GLM-5 has achieved key upgrades at the model architecture level. The parameter scale has expanded from the previous generation's 355B (activation 32B) to 744B (activation 40B), and the amount of pre-training data has increased from 23T to 28.5T, with larger computational power investments driving a significant enhancement in general intelligence capabilities. This model introduces the DeepSeek sparse attention mechanism for the first time, effectively reducing deployment costs and improving Token utilization efficiency while maintaining the long text processing effect without loss. This technical route is consistent with DeepSeek-V3/V3.2. In terms of architectural configuration, GLM-5 constructs 78 hidden layers, integrating 256 expert modules, activating 8 at a time, with activation parameters of approximately 44B, sparsity of 5.9%, and a maximum context window supporting 202K tokens. Significant Improvement in Programming Capabilities The new generation flagship model GLM-5 has performed outstandingly in the internal Claude Code evaluation. In programming development scenarios such as front-end, back-end, and long-range tasks, this model has comprehensively surpassed 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. The official claim is that the user experience in real programming environments has approached the level of Claude Opus 4.5. KNOWLEDGE ATLAS positions GLM-5 as the latest generation flagship dialogue, programming, and agent model, focusing on enhancing its processing capabilities in complex system engineering and long-range agent tasks. Agent Capabilities Achieve Optimal Performance in Open Source GLM-5 has achieved open-source SOTA in agent capabilities, ranking first in multiple evaluation benchmarks. In three tests—BrowseComp (networked retrieval and information understanding), MCP-Atlas (large-scale end-to-end tool invocation), and τ2-Bench (automatic agent tool planning and execution in complex scenarios)—GLM-5 has achieved optimal performance To achieve a breakthrough in capabilities, this model has constructed a brand new "Slime" training framework, supporting larger model architectures and more complex reinforcement learning tasks, significantly improving the efficiency of the post-training process in reinforcement learning. In addition, KNOWLEDGE ATLAS has proposed an asynchronous agent reinforcement learning algorithm, enabling the model to continuously learn from long-range interactions, effectively stimulating the deep potential of the pre-trained model. This mechanism has become one of the core technical features of GLM-5. Domestic Large Models Intensively Released During the Spring Festival The release of KNOWLEDGE ATLAS's GLM-5 has become the latest highlight in the competitive landscape of domestic AI large models during the Spring Festival. On the same evening, Minimax also launched Minimax 2.5, just over a month after the release of the previous version 2.2. This wave of releases has been heating up. DeepSeek has previously launched a new model, and products such as Alibaba's Qwen 3.5 and ByteDance's SeeDance 2.0 have also recently made their debut. Many manufacturers have chosen to concentrate their new releases during the Spring Festival window, reflecting that the competition in the domestic large model track is entering a heated stage. Currently, the detailed technical documentation for GLM-5 and Minimax 2.5 has not yet been fully disclosed, and their actual performance remains to be validated by the developer community and professional institutions