Zhipu AI has debuted GLM-5, a 744-billion parameter open-source model trained on 28.5 trillion token...
The AMW Read
The release of a 744B parameter open-weight model optimized for non-Nvidia silicon (Huawei) validates the CN/OSS challenger frame and signals a structural decoupling in the global AI compute and model landscape.
NoveltySignificance
Foundation Models · Player MapScaling LawsGeopolitics
Zhipu AI has debuted GLM-5, a 744-billion parameter open-source model trained on 28.5 trillion tokens using DeepSeek Sparse Attention. The model marks a shift from chat to agentic engineering, enabling direct creation of .xlsx and .pdf files while utilizing a new Slime RL system to cut training bottlenecks. At 1.00 dollar per million input tokens, it is 6x more affordable than Claude 4.6 and optimized for Huawei Ascend chips. This signals a strategic decoupling in global AI, prioritizing autonomous industrial automation. 🚀
