Recently, ModelBest, in collaboration with Tsinghua University and the OpenBMB open-source community, officially released and open-sourced its latest achievement in low-bit large model training—BitCPM-CANN. This is the first ternary (1.58-bit) large model to be fully end-to-end trained and open-sourced based on a domestic computing platform (Huawei Ascend).

For a long time, the physical bottleneck of memory has been a challenge for the large-scale application of large models, with memory becoming one of the most strained resources in the global AI supply chain. Against this backdrop, BitCPM-CANN adopts a quantization-aware training approach, forcing every bit to deliver maximum information density and knowledge-carrying efficiency. Meanwhile, the 6x memory optimization provided by BitCPM-CANN allows enterprises to enhance model capabilities or service density without increasing physical memory.

It is worth noting that from the most fundamental quantization operators and quantization-aware training algorithms to the complete parallel strategy and training framework, the entire training pipeline of BitCPM-CANN was natively completed on Huawei Ascend. It includes four model sizes—0.5B, 1B, 3B, and 8B—and when evaluated item-by-item against the same-sized MiniCPM-4 full-precision family, it demonstrated excellent performance. This is the first publicly available result on the Ascend platform to complete end-to-end 1.58-bit training with a full-precision comparative evaluation, and the model scale has been advanced to the 8B level in one go.

Industry insiders believe that the release and open-sourcing of BitCPM-CANN achieves a complete closed loop of domestic NPUs, domestic models, and a domestic training framework, providing a directly usable low-bit model solution for the end-side AI industry.