Chinese AI Tools
ProductsModelsIntegrationsRankingsLatest changes
Availability TrackerUse casesSubmit a toolAccount
ZHSearch

Chinese AI Tools

Independent directory for Chinese AI products. Product availability, pricing and terms can change. Verify before commercial use.

Editorial standardsClaim productUpdate infoGet featuredAdvertise

news

Huawei's openPangu 2.0 Flash and Pro Weights Are Now Public

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 2026-06-12

Verdict

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.

Huawei announced openPangu 2.0 at HDC 2026 on June 12. The original report described two highly sparse models optimized for Ascend hardware and HarmonyOS agents, with seven component groups due to open from June 30. That milestone has now partially landed: the official openPangu Hugging Face organization published openPangu-2.0-Flash and its Int8 variant on June 30, followed by openPangu-2.0-Pro and Pro-Int8 on July 30. The model repositories include downloadable safetensors and custom runtime code, so the release should no longer be described as announcement-only.

Two released model families

Both variants are now distributed through the official openPangu organization, including full-precision and Int8 repositories.

openPangu 2.0 Pro

505B total parameters with 18B activated parameters, a roughly 28:1 total-to-active ratio.

openPangu 2.0 Flash

92B total parameters with 6B activated parameters, positioned as the lighter deployment option.

Ascend and HarmonyOS positioning

Conference reports say the models are trained natively for Ascend and optimized for HarmonyOS agent tasks.

Reported inference throughput

Huawei claims per-card user throughput can reach twice that of unspecified mainstream open models. The comparison setup has not yet been published.

Reported training gains

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.

What is now public

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.

Model assets

Flash, Flash-Int8, Pro and Pro-Int8 repositories are public with safetensors and custom runtime code.

Training stack

The original announcement also promised pre-training code, post-training code and training operators; verify these components independently from the weight repositories.

What remains unknown

The weight release resolves availability, but developers still need license, hardware and reproducibility review before deployment.

License

The repositories use a custom model license rather than a standard OSI software license; review its commercial and redistribution clauses directly.

Hardware portability

Ascend-native optimization is central to the launch; support and performance on CUDA or other accelerators remain unconfirmed.

Independent evaluation

Weights are public, but vendor benchmarks still need independent reproduction on disclosed hardware and runtime settings.

Sources

LocalLLaMA community postHDC 2026 openPangu 2.0 reportHuawei open-source programopenPangu Hugging Face organization

Next actions

  • - Review the repository license and deployment code before commercial or cross-hardware use.
  • - Do not estimate local VRAM requirements from activated parameters alone; wait for weight precision, expert layout and runtime documentation.