BitCPM-CANN

First publicly reported 1.58-bit ternary quantization-aware training at 8B scale on Huawei Ascend NPU (CANN stack). Four sizes released: 0.5B, 1B, 3B, 8B. The 8B variant retains 95.7% of FP accuracy, achieves ~6× memory reduction, and imposes only ~5% throughput overhead vs full-precision training.

Significant domestic-chip co-design milestone — demonstrates a fully China-domestic training stack for bitnet-class models, paralleling ByteDance's recently disclosed custom-CPU effort and DeepSeek-V4-Pro's Ascend 950PR pre-training. By the MiniCPM team.