面壁智能放大招！开源全尺寸BitCPM-CANN：国产算力首次跑通1.58-bit训练，推理显存省5/6

Mianbi Intelligence has unveiled the open-source BitCPM-CANN model, achieving 1.58-bit training without relying on expensive hardware. This breakthrough reduces inference memory usage by 5/6, marking a significant advancement for domestic AI capabilities.

Key Points

• BitCPM-CANN achieves 1.58-bit training, a first for domestic AI.
• Inference memory usage is reduced by 5/6, enhancing efficiency.
• No reliance on expensive hardware, making AI more accessible.
• This development opens new avenues for domestic AI innovation.
• Mianbi Intelligence leads the charge in affordable AI solutions.