Huawei Open-Sources openPangu-2.0 Pretrain, SFT and RL Training Code on Ascend, Pandaily reports
Reported by Pandaily · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Huawei open-sourced openPangu-2.0 pretrain, SFT and post-training RL code for its Ascend-native MoE stack on September 28, 2026, complementing earlier Pro (505B) and Flash (92B) weight releases, Pandaily reported.
China context
- Original name
- 华为
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can access Ascend-optimized training code for large MoE models, potentially enabling training on Huawei hardware.
- For investors
- The release may strengthen Huawei's position in the AI infrastructure market by expanding its software ecosystem.
On September 28, 2026, Huawei released the training code for openPangu-2.0, covering pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL) post-training. The code targets Huawei's Ascend AI hardware and complements previously released model weights: openPangu-2.0 Pro with 505 billion parameters and openPangu-2.0 Flash with 92 billion parameters. The release is part of Huawei's open-source efforts around its Pangu large language model family.
The release includes code for the full training pipeline—pretraining, SFT, and RL—specifically optimized for Ascend NPUs. This enables developers to reproduce or extend training on Huawei's hardware stack. The MoE (Mixture of Experts) architecture is noted, with two model sizes: 505B (Pro) and 92B (Flash).
Developers outside China gain access to Huawei's Ascend-optimized training code, potentially lowering the barrier to train large MoE models on non-NVIDIA hardware. This could affect hardware and software choices for organizations seeking alternatives to CUDA-based ecosystems.
The open-sourcing of training code may reduce development costs for enterprises using Ascend hardware and could increase adoption of Huawei's AI stack in markets where Ascend is available.
A specific signal to watch is whether independent developers successfully reproduce training runs using the released code on Ascend hardware, and whether the code is adopted in projects outside Huawei's ecosystem.