Huawei Opens the openPangu-2.0 Training Stack for Ascend

Huawei’s September 28, 2026 open-source release adds code for openPangu-2.0 pretraining, supervised fine-tuning (SFT) and post-training reinforcement learning (RL). The key difference is what developers can work with: earlier Pro and Flash releases made model weights and inference resources available, while the new Ascend-oriented projects cover stages for training and adapting models. 1227

What Huawei released

The newly announced openPangu-2.0-Training project supports pretraining and SFT, while openPangu-2.0-RL focuses on reinforcement-learning post-training. Both are designed for Huawei’s Ascend training ecosystem. 1

That makes this a training-code release, not simply another model checkpoint. The announcement describes the code projects and their focus; it does not, by itself, establish the compute, data or configuration needed to reproduce the original training runs. 1

Pro and Flash at a glance

The published model cards describe both versions as mixture-of-experts (MoE) models. Their reported specifications are: 1927

Model	Total parameters	Active parameters per token	Context window	Reported pretraining tokens
openPangu-2.0-Pro	About 505B	About 18B	512K tokens	About 34T
openPangu-2.0-Flash	About 92B	About 6B	512K tokens	About 34T

These are the reported figures for each model. The training-token figure is about 34 trillion for each model, not a combined total. 1927

What the model cards say about post-training

The Pro and Flash cards describe a unified SFT stage intended to bring together “slow” and “fast” thinking, followed by multiple specialist RL stages and online policy distillation (OPD). These model-card descriptions explain the models’ reported post-training approach; the September release separately makes training and RL code available. 19271

Architecture features described in the cards

The cards describe a combination of multi-head latent attention (MLA) and DSA/SWA attention layers in a 1:2 ratio. Sliding-window attention (SWA) handles local context, while the sparse DSA layers aggregate longer-range information. 1927

They also list a four-stream mHC architecture, a three-head multi-token prediction (MTP) module and the Muon optimizer. These are model-card-reported design features, rather than capabilities newly announced as part of the September code release. 1927

What changes for Ascend developers

Previously released weights and inference code let developers run the published Pro or Flash models. The new projects extend the available tooling to pretraining, SFT and RL post-training, giving Ascend developers code to explore those parts of the model lifecycle as well. The sources identify the projects’ broad focus, but do not detail every supported workflow or its hardware requirements. 1227