Huawei released the weights, inference code, and technical report for openPangu-2.0-Pro on July 31, a 505-billion-parameter language model that completed its full pretraining run on Huawei’s Ascend 910B NPUs with no Nvidia hardware involved at any stage. According to TechTimes, no prior frontier-scale open-weight model at 500 billion parameters or above has credibly made that claim.
The model is live on GitCode’s Ascend Tribe community and accessible through Huawei Cloud ModelArts Studio, according to AIBase. Community efforts to run the model on Nvidia GPUs via standard format conversions are also underway.
Architecture
openPangu-2.0-Pro uses a Mixture of Experts (MoE) architecture with approximately 18 billion parameters activated per token during inference, according to AIBase. It supports a 512K context length and was trained on approximately 34 trillion tokens. Post-training included supervised fine-tuning, multiple rounds of specialized reinforcement learning, and capability fusion through online distillation.
The model is part of Huawei’s broader openPangu 2.0 open-source plan. At HDC 2026 in June, Huawei announced it would gradually open-source seven key components starting June 30, including pre-training code, post-training code, and training operators, according to AIBase. A smaller 92-billion-parameter openPangu-2.0-Flash model was released first.
The Supply Chain Behind “No Nvidia”
The Ascend 910B NPUs that trained openPangu-2.0-Pro are manufactured primarily on SMIC’s N+2 process, a 7-nanometer-class node that uses deep ultraviolet (DUV) lithography, not the extreme ultraviolet (EUV) systems that TSMC, Samsung, and Intel use for leading-edge nodes. ASML holds a near-monopoly on EUV systems, and export agreements bar shipments to China, according to TechTimes.
But the immediate complication lies in what was inside the Ascend 910B chips at the time of production. TechInsights teardowns of Ascend 910B and 910C samples found that essentially every chip examined contained dies fabricated at TSMC’s 7nm node, not SMIC’s, according to TechTimes. Huawei had acquired approximately 2.9 million TSMC-fabricated 7nm dies through Sophgo, a Cayman Islands-registered chip design firm. U.S. authorities determined the scheme violated export controls. TSMC subsequently faced a potential penalty of $1 billion or more under a Department of Commerce investigation.
Samsung HBM stacks were also found alongside the TSMC compute dies in the teardowns. China stockpiled approximately 13 million Samsung HBM stacks before export controls on HBM exports tightened, according to TechTimes.
The Real Test Comes Next
As of early 2026, the TSMC die bank is effectively exhausted, according to TechTimes. Future Ascend production depends entirely on SMIC-fabricated wafers and domestically packaged high-bandwidth memory from CXMT (ChangXin Memory Technologies), which has not yet reached the volume or yield reliability of Samsung or SK Hynix.
openPangu-2.0-Pro demonstrates that the Ascend software and architecture stack can sustain frontier-scale training. It does not yet demonstrate that a fully domestic hardware stack can do the same, because that hardware did not train this model. The next Ascend-based frontier training run will be the real test of China’s ability to train large models independently of foreign fabrication.