Huawei has released the weights, inference code and technical report for its 505B open-weight AI model, offering researchers the first public blueprint for frontier-scale Ascend-native training without Nvidia GPUs.
Huawei has released the weights, inference code and technical report for openPangu-2.0-Pro, a 505-billion-parameter open-weight language model that it says completed its entire pretraining run on Ascend 910B NPUs without using Nvidia GPUs. If validated, it would be the first publicly released open-weight model above 500 billion parameters to achieve that milestone.
The model is now available through GitCode’s Ascend Tribe community and Huawei Cloud ModelArts Studio, allowing developers to download the weights immediately. Community efforts to convert the model for Nvidia GPU inference are also underway, broadening access beyond Huawei’s hardware ecosystem.
Built on a Mixture of Experts (MoE) architecture with 18 billion active parameters per token, openPangu-2.0-Pro features a 512,000-token context window and was pretrained on approximately 34 trillion tokens. Huawei also details its use of the Muon optimiser, Multi-head Latent Attention, Decoupled Sparse Attention, Sliding Window Attention and a three-stage post-training pipeline featuring Online Policy Distillation (OPD).
However, the release does not establish a fully domestic Chinese AI hardware supply chain. The article notes that earlier Ascend chips reportedly contained TSMC-fabricated 7nm compute dies and Samsung HBM memory, while future production is expected to rely on SMIC-manufactured chips and CXMT memory. Huawei’s reported performance claims also remain unverified by independent benchmark organisations.
The release nonetheless provides researchers and sovereign AI programmes with one of the most detailed public references yet for building frontier-scale AI models on an Ascend-native software stack.
