Core Insights - The launch of GLM-5 marks a significant advancement in domestic AI models, focusing on programming and agent capabilities, and is claimed to achieve optimal performance in the open-source domain [1][5] Group 1: Model Specifications - GLM-5's parameter scale has increased from 355 billion to 744 billion, with activation parameters rising from 32 billion to 40 billion [2] - The pre-training data volume has expanded from 23 terabytes to 28.5 terabytes, enhancing general intelligence capabilities [2] - The model incorporates a new sparse attention mechanism called DeepSeek, which reduces deployment costs while maintaining long text processing efficiency [2] Group 2: Performance Enhancements - In internal evaluations, GLM-5 outperformed its predecessor GLM-4.7 by over 20% in various programming scenarios, including front-end, back-end, and long-range tasks [3] - The model is capable of autonomously completing complex system engineering tasks with minimal human intervention, achieving a programming experience comparable to Claude Opus 4.5 [3] Group 3: Agent Capabilities - GLM-5 has achieved state-of-the-art (SOTA) performance in agent capabilities, ranking first in multiple evaluation benchmarks [4] - The model features a new training framework called "Slime," which enhances the efficiency of reinforcement learning tasks and supports larger model architectures [4] - An asynchronous reinforcement learning algorithm has been introduced, allowing the model to learn continuously from long-range interactions [4] Group 4: Industry Context - The release of GLM-5 is part of a broader trend of domestic AI model launches during the Spring Festival period, indicating intensified competition in the sector [5][6] - Other companies, such as Minimax, Alibaba, and ByteDance, have also recently introduced new models, reflecting the competitive landscape of domestic AI development [6]
智谱发布新一代旗舰模型GLM-5，重点提升编程与智能体能力