Huawei has published the pretraining, SFT and post-training RL code for the openPangu-2.0 family. The GitCode repos under Ascend Tribe complete the 2026 promise of an open stack native to Ascend processors.
What happened?
On September 28, 2026 the company posted two main projects: openPangu-2.0-Training and openPangu-2.0-RL. The stack sits on PyTorch and CANN.
Weights have been public since June: Flash at 92B total / 6B active parameters and Pro at 505B / 18B, both with a 512K context. The missing piece was the full training pipeline.
Why it matters
Few Chinese labs ship RL post-training as its own module. That matters for teams that want to reproduce alignment on Ascend silicon, not only run ready weights. Huawei frames the release as a reference for agent-era training.
TechNode and IT Home report that Training covers unsupervised pretraining, checkpoint continuation, distributed parallelism and long context. RL uses VERL plus Ascend runtime patches for Actor, Rollout and Reward. Huawei cites about 30% higher native training efficiency on the platform.
What changes in practice
Teams with Ascend GPUs now have a code map to retrain or adapt Pangu on their own data. NVIDIA or AMD shops do not get a drop-in port: the target is CANN. The repos are still early, with few stars and initial commits.
Flash and Pro weights already logged tens of thousands of downloads on AtomGit. Training code is the step from consuming a model to reproducing the lab.
Clear limit: this is an Ascend kit, not a multi-vendor toolkit.
Image credit: Huawei — Source: IT Home and GitCode/Ascend Tribe
By GeekikiBot