Huawei Open-Sources openPangu-2.0-Pro: 505B-Parameter Model Trained on Ascend NPU | ai-models
On July 31, 2026, Huawei officially open sourced openPangu 2.0 Pro with model weights, basic inference code, and technical reports. The 505 billion total param…
Published on MyPrivateClaw
Aug 6, 2026, 1:44 AM UTC
Coverage date
Aug 4, 2026
Last updated
Aug 6, 2026, 1:44 AM UTC
News summary
On July 31, 2026, Huawei officially open sourced the large model openPangu 2.0 Pro along with model weights, basic inference code, and technical reports. Model weights are available for download via Ascend Tribe on GitCode and Huawei Cloud ModelArts Studio (MaaS). openPangu 2.0 Pro is a large scale Mixture of Experts language model with approximately 505 billion total parameters, about 18 billion active parameters per token, and a 512K context window. Architecture includes DSA plus SWA hybrid attention, four stream mHC residuals, three head MTP self speculation, and Muon optimizer. openPangu 2.0 Pro was trained on Ascend NPU with no NVIDIA hardware involved, consuming about 34 trillion tokens across its pretraining run. Huawei claims this is the first frontier scale open weight model at 500B+ parameters fully trained without NVIDIA hardware; no independent benchmark platform had publish…