Bingbi AI Open-Sources BitCPM-CANN: 1.58-bit Training Achievable on Domestic Compute

Bingbi AI open-sourced BitCPM-CANN, a full training pipeline for 1.58-bit model training on domestic AI accelerators.

The release includes code, checkpoints, and benchmarks, showing the method can run on platforms such as Huawei Ascend.

By cutting memory use, the approach could ease training bottlenecks and lower costs for large-model development.

It also highlights a path toward less dependence on restricted NVIDIA hardware in the Chinese AI ecosystem.