Mianbi Intelligence, Tsinghua University, and the OpenBMB community have released BitCPM-CANN, the first native end-to-end 1.58-bit (ternary) large language model training system built on Huawei's Ascend NPU. The system's "ternary quantization" compresses model weights into just three states: -1, 0, and 1. This slashes memory usage by roughly 6x and turns power-hungry floating-point multiplications into simple additions, creating a complete pipeline from low-level quantization operators to distributed training on Ascend 910B.
https://twitter.com/OpenBMB/status/2057816337880355220
The team is releasing four model sizes: 0.5B, 1B, 3B, and 8B, and benchmarked them across 11 tests against the full-precision MiniCPM4. The 1B, 3B, and 8B ternary versions retain 97.1%, 97.2%, and 95.7% of their full-precision counterparts' average performance, respectively. The 8B version's slightly lower retention rate is due to its handling of more complex knowledge, but its absolute score of 77.84 still far exceeds the ternary 3B's 72.32, keeping high intelligence while lowering deployment barriers. The 0.5B version retains 90.1%.
Training overhead is minimal: on Ascend 910B, quantized training achieves 148 TFLOP/s versus 155 TFLOP/s for full precision — an extra compute cost of just ~4.5%. This initial release is a pseudo-quantized version where weights are constrained to ternary values but stored as floats. That means you can load it like any regular model using Hugging Face's Transformers library, no special quantization library needed, making it much easier to deploy on phones, PCs, and other edge devices.