GLM-5: Targeting complex systems engineering and long-horizon agentic tasks (z.ai)

The launch of GLM-5 marks a significant advancement in AI and machine learning, specifically focusing on complex systems engineering and long-horizon agentic tasks. Scaling up from 355 billion parameters in GLM-4.5 to 744 billion in GLM-5, along with an increase in pre-training data from 23 trillion to 28.5 trillion tokens, reflects a strategic push to enhance the effectiveness of Artificial General Intelligence (AGI). The integration of DeepSeek Sparse Attention (DSA) not only reduces deployment costs but also maintains robust long-context capabilities, making GLM-5 highly efficient. Moreover, with the novel asynchronous reinforcement learning infrastructure dubbed "slime," GLM-5 achieves significant improvements in training throughput and fine-tuning capabilities. Its performance across various benchmarks, including being ranked #1 in Vending Bench 2 for long-term operational tasks, demonstrates its superior reasoning, coding, and planning abilities, nearing the performance of proprietary models like Claude Opus 4.5. Open-sourced on platforms like Hugging Face and ModelScope under the MIT License, GLM-5 is readily accessible for developers, signaling a pivotal move towards making advanced AI capabilities available for diverse applications in the industry.