GLM-5 represents a significant upgrade in artificial intelligence, focusing on complex systems engineering and long-horizon tasks. It increases its parameters from 355 billion in GLM-4.5 to 744 billion in GLM-5, alongside a spike in pre-training data from 23 trillion to 28.5 trillion tokens. The introduction of DeepSeek Sparse Attention (DSA) helps reduce deployment costs while maintaining long-context capacity, making the model more efficient. A new asynchronous reinforcement learning infrastructure, called slime, boosts training efficiency and throughput, allowing for detailed post-training adjustments. In benchmark tests, GLM-5 outperforms its predecessor and ranks highly against other models in reasoning, coding, and agentic tasks. Notably, in Vending Bench 2, GLM-5 achieved a final account balance of $4,432 in a simulated vending machine operation, coming close to Claude Opus 4.5. The model is open-sourced on platforms like Hugging Face and ModelScope, with MIT License model weights. It’s accessible via developer platforms and can be trialed for free on Z.ai. GLM-5’s capabilities extend beyond basic functions; it can generate various document formats, supporting tasks like creating PRDs and reports. The official application is introducing an Agent mode, enhancing multi-turn collaboration and making it easier to produce professional documents.