The 1.58-bit ternary model run on Ascend 910B, with memory reduced to one-sixth of BF16, can still maintain over 95% accuracy. Edge deployment is finally feasible.

BitCPM-CANN is jointly released by ModelBest, Tsinghua, and the OpenBMB community, and is the world's first open-source 1.58-bit ternary model trained on the Ascend 910B NPU. It adopts tri-state weight ultra-low bit quantization, reducing memory usage by approximately 6 times compared to BF16, making it suitable for edge devices such as smartphones, computers, and vehicles. The training stack is fully native on Ascend, with the 0.5B–8B family achieving 95–97% full-precision performance on multiple benchmarks, making deployment more feasible.