GLM-5 was trained entirely on Huawei chips (glm5.net)

Zhipu AI has announced the launch of GLM-5, its fifth-generation large language model, which boasts an impressive 745 billion parameters, with a Mixture of Experts architecture featuring 44 billion active parameters per inference. Trained entirely on Huawei Ascend chips using the MindSpore framework, GLM-5 achieves full independence from US hardware, marking a significant step in China's development of a self-reliant AI infrastructure. This model excels in various domains including creative writing, advanced coding capabilities, multi-step reasoning, agentic intelligence, and efficiently handles long-context processing with a capacity of up to 200K tokens. The significance of GLM-5 within the AI/ML community lies in its potential to rival established models like OpenAI's GPT-5 and Anthropic's Claude, particularly due to its competitive performance and cost-efficiency. With an expected pricing model significantly lower than its counterparts—approximately $0.11 per million tokens—GLM-5 is set to make frontier-level AI technology more accessible to a wider audience. As Zhipu AI has a strong track record of open-sourcing its models, GLM-5 is anticipated to be released under the MIT license, further promoting community engagement and innovation in AI development.