Huawei has released the source code for the pretraining, Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL) processes for the openPangu-2.0 model. This release occurred on September 28, 2026, according to information from TMT Post and coverage related to the training stack optimized for Ascend.
The official branding is used as Huawei / openPangu. Previously, the weights for openPangu-2.0-Pro and openPangu-2.0-Flash were released on the Hugging Face platform; however, the current release pertains specifically to the training code path that underpins these MoE models trained on Ascend, rather than releasing new parameters.
The model cards on Hugging Face describe openPangu as Huawei's brand of open artificial intelligence models designed for training and inference on Ascend. The openPangu-2.0-Pro model features approximately 505 billion total parameters with about 18 billion active per token, a 512K context window, and a training budget of around 34 trillion tokens. The openPangu-2.0-Flash model has about 92 billion total parameters, approximately 6 billion active, the same 512K context size, and a comparable budget of ~34 trillion tokens.
Additional details following the training of both models mention combined fast/slow SFT, multi-specialized RL, and Online Distillation (OPD). Architectural features common to Pro and Flash include multi-head latent attention, DSA-plus-SWA layered mix (approximately 1:2), an mHC residual topology with four branches, three-head multi-token prediction, and training using the Muon optimizer.
The September 28th release provides the pretraining, SFT, and RL components as Ascend-specific tools, allowing developers to reproduce and extend this stack instead of merely loading weights for inference. It is important to distinguish between open weights, inference code, and this training code set: the weights were already publicly available; the news is that Huawei is opening up the openPangu-2.0 training pipeline on the Ascend side for the pretrain/SFT/RL components.
For teams working on Ascend, the concrete outcome is an OSS training stack corresponding to the confirmed Pro family models (505B / ~18B active, 512K) and Flash (92B / ~6B active, 512K).
