What does training without NVIDIA mean?

GLM-5 is the first frontier-class model trained entirely on Huawei Ascend chips, proving viability of Chinese chips for large-scale training amid US export controls.

What does 744B total but 40B active mean?

This is the MoE architecture: the model has 744B total parameters but only activates 40B per inference, keeping the knowledge capacity of a large model with the inference cost of a 40B model.

How was the record-low hallucination achieved?

Through 28.5T tokens of high-quality pretraining, DeepSeek Sparse Attention for precise attention mechanisms, and 'slime' asynchronous RL alignment training.

— AI DAILY

Sources

Zhipu AI
VentureBeat
The Decoder