Chinese AI developer Zhipu AI, with support from ByteDance, has released the technical report and open-source code for GLM-4.5, a Mixture-of-Experts large language model aimed at reasoning, coding and autonomous-agent tasks. The system contains 355 billion parameters, with 32 billion activated during inference, and was trained on 23 trillion tokens using a multi-stage regimen that combines expert iteration, self-distillation and reinforcement learning. According to the report, GLM-4.5 achieves 70.1 % on the TAU-Bench reasoning test, 91.0 % on the AIME 24 math benchmark and 64.2 % on the SWE-bench Verified software-engineering suite. The scores place the model third overall across 12 widely used evaluation sets and second on dedicated agentic benchmarks, while using fewer active parameters than many competing systems. Zhipu AI also released a smaller 106-billion-parameter version, GLM-4.5-Air, as well as GLM-4.5V, a vision-language variant that reports state-of-the-art results on 42 image, video and document-understanding tasks. The models and accompanying code are available on Hugging Face and GitHub under open-source licenses to encourage further research and commercial adoption.