BitCPM-CANN Released as First 1.58-bit Open Model Fully Trained on Huawei Ascend 910B NPU
首个基于华为昇腾910B NPU全栈训练的1.58比特开源大模型BitCPM-CANN发布
ModelBest, Tsinghua University, and OpenBMB released BitCPM-CANN, a 0.5B-8B open model family trained natively on Huawei Ascend 910B NPUs with 1.58-bit ternary weights, cutting memory use by about 6x versus BF16 while retaining 95-97% of full-precision benchmark performance.
Why it matters: ModelBest, Tsinghua and OpenBMB released BitCPM-CANN, trained entirely on Huawei Ascend 910B, where 1.58-bit quantization saves about 6x memory over BF16 while benchmarks hold 95-97% of full-precision scores.