Bingbi AI Open-Sources BitCPM-CANN: 1.58-bit Training Achievable on Domestic Compute
Bingbi AI has released BitCPM-CANN, an open-source training framework designed for efficient model training. This framework enables 1.58-bit model training, which significantly reduces memory requirements for inference, potentially up to six times less than traditional full-precision methods. The development aims to make advanced AI training more accessible, particularly on domestic compute hardware. AI
IMPACT Enables more efficient AI model training and inference, potentially lowering hardware requirements and costs.