← Back to feed

A team from ModelBest, Tsinghua University, and the OpenBMB community has open-sourced BitCPM-CANN, the first 1.58-bit ternary language model trained entirely on Huawei Ascend 910B NPUs. The model family covers 0.5B to 8B parameters and is available on Hugging Face and ModelScope under an open license.

The technical claim is specific: this is not a post-training quantization or a model ported to Ascend after training on NVIDIA hardware. The entire pipeline — quantization-aware training (QAT), straight-through estimation (STE), low-bit operators, and the full framework stack — was natively built and executed on Ascend 910B from the start.

What 1.58-Bit Weights Mean

In 1.58-bit (ternary) quantization, every weight takes one of three values: -1, 0, or +1. This allows the model to replace expensive multiply-accumulate operations with additions and subtractions, dramatically cutting memory bandwidth requirements.

The practical result: roughly 6x lower memory footprint compared to the BF16 baseline. An 8B parameter model that would require ~16 GB in full precision fits within 3-4 GB in 1.58-bit format — running on mobile hardware, in-car compute, or edge industrial devices without modification.

Benchmark retention across the 11-task evaluation suite runs at 95-97% for the 1B-8B range compared to full-precision MiniCPM4, the baseline model family BitCPM-CANN derives from.

The Harder Claim: Ascend-Native Training

Running inference on Huawei hardware is not new. Running a complete QAT pipeline on it at this scale is. The team had to build quantization operators and algorithms, a low-bit training framework, and reproducible training scripts that functioned reliably across the Ascend 910B environment.

This matters because the training stack determines what runs in production. A model trained on NVIDIA H100s and then deployed on Ascend typically carries compatibility debt — performance gaps, missing kernel coverage, edge cases in the quantization path. BitCPM-CANN’s entire quantization path was developed on and for Ascend, which means the deployment is native, not adapted.

The team reports the model has been verified across all 11 benchmark tasks at the 0.5B through 8B sizes, with open-source code covering the full pipeline from research to deployment.

Context: Compute Constraints as a Design Driver

The release aligns with a broader pattern in Chinese AI lab output: hardware limitations producing algorithmic efficiency that NVIDIA-first labs have less incentive to develop.

DeepSeek’s V4 launch in April 2026 noted that adapting its training stack to Huawei Ascend silicon was one of the primary technical challenges its new funding would address. The BitCPM-CANN work from ModelBest and Tsinghua represents a parallel track — not adapting existing models to Ascend, but building new ones natively on it, with 1.58-bit quantization as the efficiency mechanism that makes the hardware constraints a feature rather than a liability.

The OpenBMB community frames this explicitly: “When hardware costs rise, the winning model is not merely the one that scores higher in a chart, but the one that can be trained, reproduced, deployed, and improved under real constraints.”

Edge Deployment Potential

At 1.58 bits, the 8B model fits within typical automotive compute envelopes without custom silicon. The team demonstrates that with MoE scaling, models approaching 100B total parameters could fit on terminal hardware at this bit depth.

That edge deployment profile makes BitCPM-CANN relevant beyond the sovereignty narrative. Any deployment context where cloud inference latency is unacceptable, per-token costs matter at scale, or hardware selection is constrained — industrial IoT, automotive, mobile AI — has a reason to look at 1.58-bit as a practical tier, not just a research result.