Guan Cha Zhe Wang·2026-02-12 02:05Core Insights - Zhipu AI has launched and open-sourced GLM-5, with multiple domestic chip manufacturers completing Day-0 adaptation for the model, ensuring compatibility and stable operation from the first day of release [1][2] Group 1: GLM-5 Launch and Adaptation - Zhipu AI's GLM-5 has been adapted by major domestic chip platforms including Huawei Ascend, Moore Threads, Cambricon, Kunlun, Muxi, Suiruan, and Haiguang, achieving high throughput and low latency on domestic computing clusters [2] - Haiguang's DCU team collaborated closely with Zhipu AI to optimize the underlying operators and hardware acceleration, enabling GLM-5 to run stably on Haiguang DCU with high throughput and low latency [1] - Moore Threads completed full-process adaptation and verification on its flagship AI training and inference GPU MTT S5000, leveraging its MUSA architecture to enhance model inference performance while reducing memory usage [1] Group 2: Model Performance and Features - GLM-5 has expanded its parameter scale from 355 billion (activated 32 billion) to 744 billion (activated 40 billion) and increased pre-training data from 23 terabytes to 28.5 terabytes, significantly enhancing its general intelligence capabilities [2] - The model incorporates a new "Slime" asynchronous reinforcement learning framework, supporting larger model scales and more complex tasks, and utilizes DeepSeek Sparse Attention to maintain long text performance while reducing deployment costs and improving token efficiency [2] - A month prior, Zhipu AI released the GLM-Image model, which employs a hybrid architecture of autoregressive and diffusion decoders, marking a significant exploration in cognitive generative technology [2]