news
The HDC 2026 announcement has moved into public release: Flash weights appeared on June 30 and Pro weights on July 30 in the official openPangu Hugging Face organization.
Published
The public weights make openPangu 2.0 testable, but practical openness still depends on its custom model license, Ascend-first runtime requirements and independent reproducible evaluation.
Both variants are now distributed through the official openPangu organization, including full-precision and Int8 repositories.
505B total parameters with 18B activated parameters, a roughly 28:1 total-to-active ratio.
92B total parameters with 6B activated parameters, positioned as the lighter deployment option.
Conference reports say the models are trained natively for Ascend and optimized for HarmonyOS agent tasks.
Huawei claims per-card user throughput can reach twice that of unspecified mainstream open models. The comparison setup has not yet been published.
The Reddit summary attributes 30% higher SuperPoD training efficiency, 50% higher 512K-sequence throughput and more than 99% training consistency to the presentation; these details need confirmation in the technical report.
The official distribution channel now exposes model weights, configuration and custom model code for Flash and Pro. Training-stack completeness should still be verified separately rather than inferred from model repositories alone.
Flash, Flash-Int8, Pro and Pro-Int8 repositories are public with safetensors and custom runtime code.
The original announcement also promised pre-training code, post-training code and training operators; verify these components independently from the weight repositories.
The weight release resolves availability, but developers still need license, hardware and reproducibility review before deployment.
The repositories use a custom model license rather than a standard OSI software license; review its commercial and redistribution clauses directly.
Ascend-native optimization is central to the launch; support and performance on CUDA or other accelerators remain unconfirmed.
Weights are public, but vendor benchmarks still need independent reproduction on disclosed hardware and runtime settings.