Mingshi Intelligence with Tsinghua University and OpenBMB open-sourced China's first large model trained on Huawei Ascend, BitCPM-CANN. It uses a 1.58-bit ternary representation, achieving breakthroughs in low-bit training. The model comes in 0.5B to 8B sizes, offering six times memory efficiency in inference, enabling an 8B model to run on flagship smartphones. All weights are open-sourced on HuggingFace and ModelScope.