Model Best Open-Sources BitCPM-CANN: 1.58-bit Training Achievable on Domestic Compute

Model Best has launched BitCPM-CANN, an open-source training framework that enables training models at 1.58-bit precision on local AI accelerators. This framework demonstrates a reduction in memory requirements during inference by up to six times compared to full precision training. This development comes at a critical time, with HBM memory prices increasing by 165% year-on-year, making memory bandwidth one of the most expensive resources for model training.