China's First Ternary LLM BitCPM-CANN Runs on Huawei Ascend

MiniCPM team, together with Tsinghua University and OpenBMB, released BitCPM-CANN—China's first ternary (1.58-bit) large language model trained entirely on Huawei's Ascend AI chips. It comes in four sizes: 0.5B, 1B, 3B, and 8B parameters, and performs well compared to the full-precision MiniCPM4 models.

Compared to traditional BF16 precision, BitCPM-CANN frees up about 6x more memory during inference while keeping 90% to 97.2% of the model's capability. That means an 8B model can run smoothly on current flagship phones, making big AI models more accessible for mobile devices.

The team built a complete low-bit training setup using MindSpeed × Megatron-LM, supporting 32K long sequences, parallel strategies, and fused operators. All future low-bit training work on Ascend can now use this common infrastructure. The full model weights are open-sourced on HuggingFace and ModelScope.