OpenBMB Launches BitNet Models for Huawei Ascend Hardware

OpenBMB has released three BitNet-architecture models — BitCPM4-CANN-1B, -3B, and -8B — explicitly optimized for Huawei's CANN (Compute Architecture for Neural Networks) stack, marking the first publicly released BitNet models targeting Ascend GPU infrastructure rather than NVIDIA CUDA. BitNet models use 1-bit weight quantization, dramatically reducing memory bandwidth and compute requirements at inference time. By targeting CANN, OpenBMB is positioning these models to run efficiently on Huawei's Ascend 910 and related chips — hardware that Chinese labs and enterprises increasingly rely on as NVIDIA export restrictions tighten. Essentially: (OpenBMB, Huawei) are jointly signaling that China's AI stack is maturing end-to-end, from hardware to open-weight model releases. - All three models are available on Hugging Face, but community benchmarking is gated on llama.cpp adding full CANN backend support, which hasn't shipped yet. - The CANN-specific optimization means these models are unlikely to run efficiently on standard CUDA setups, narrowing the immediate addressable user base outside China. - BitNet's 1-bit architecture is still considered experimental for production use, so the primary audience is researchers and hardware developers validating the Ascend inference stack. The release is less about model capability and more about infrastructure signaling: Chinese research groups are now shipping the full software layer on top of domestic silicon.