Huawei’s openPangu 2.0 releases pretraining, SFT and reinforcement-learning code for Ascend AI chips, expanding its open-source ecosystem for large language model development.
Huawei has released openPangu 2.0, a major update to its open-source AI ecosystem that makes key stages of large language model development publicly available. The release includes pretraining, supervised fine-tuning (SFT) and reinforcement-learning (RL) training code, giving developers access to a more complete workflow for building and adapting AI models on Huawei’s Ascend AI hardware.
Unlike earlier releases that focused mainly on model weights, openPangu 2.0 exposes the training infrastructure itself. Developers can now access the code used for pretraining foundation models, instruction tuning through SFT and RL-based post-training, allowing researchers to reproduce and customise different stages of the training pipeline.
The release is closely tied to Huawei’s Ascend AI ecosystem, with the software optimised for CANN, MindSpore and Ascend accelerators. This enables developers working on Huawei’s hardware platform to train, fine-tune and deploy large language models using an integrated open-source software stack rather than relying on separate proprietary components.
Huawei’s openPangu family includes models across multiple scales, ranging from smaller developer-focused models to larger Mixture-of-Experts (MoE) variants intended for more demanding AI workloads. The broader openPangu roadmap also covers industry-focused models alongside tools for model training, optimisation and deployment within the Ascend ecosystem.
By releasing openPangu 2.0, Huawei is expanding the amount of its AI development pipeline available as open source, moving beyond model distribution towards reproducible training workflows. The update provides researchers and developers with a more complete set of tools for experimenting with large language models while strengthening Huawei’s growing open-source AI ecosystem around Ascend hardware.
