Author | Li Shuiqing, Editor | Xin Yuan
According to a report by Zhidx on September 28, Huawei announced that the pre-training, SFT (Supervised Fine-Tuning) code, and post-training RL (Reinforcement Learning) code for the open-source version of the Pangu Large Model 2.0, openPangu-2.0, have been officially open-sourced and released.
Open-source repository: https://gitcode.com/ascend-tribe
The openPangu Training Framework is a unified training framework designed for large-scale foundation model training, supporting the training and development of large language models, vision-language models, and multimodal models. The framework provides integrated training capabilities ranging from large-scale pre-training and supervised fine-tuning to continuous model optimization, and supports multi-dimensional distributed parallelism, efficient data pipelines, mixed-precision training, custom operator optimization, and highly available training.
The repository primarily provides the following training capabilities: In terms of pre-training, it supports training foundation models from scratch based on large-scale unsupervised corpora, continuous training based on existing checkpoints, large-scale distributed training, and long-context training. In terms of supervised fine-tuning, it supports instruction alignment and capability enhancement based on pre-trained models, fine-tuning with instruction data, multi-turn dialogue data, and domain-specific data, as well as parameter-efficient fine-tuning and multi-task training.
openPangu-2.0-RL is a reinforcement learning acceleration framework tailored for Huawei Ascend. Using the open-source reinforcement learning framework VERL as its training orchestration core, it completes the collaborative training of Actor (policy model), Rollout (sample generation), and Reward (reward model) on Ascend clusters through runtime patches and Ascend affinity optimizations.
For the post-training of openPangu using GSPO (Group Sequence Policy Optimization) / GRPO (Group Relative Policy Optimization), typical use cases include single-turn Rollout for mathematical reasoning, and multi-turn Rollout (ReAct, Reasoning and Acting framework) for reasoning-oriented code generation. Previously, on June 12, Richard Yu, Managing Director of Huawei and Chairman of the Terminal BG, officially released openPangu-2.0 for the first time at the Huawei Developer Conference, proposing to "build the best Pangu large model in the world." Subsequently, on June 30, openPangu-2.0-Flash, with a total of 92 billion parameters and 6 billion active parameters, was open-sourced; on July 31, openPangu-2.0-Pro, with a total of 505 billion parameters and 18 billion active parameters, was open-sourced.
According to its technical report, the openPangu-2.0 series integrates self-developed hardware cores, full-system architecture, and integrated training-inference agent reinforcement learning algorithms, specifically designed for long-context agent tasks. Meanwhile, the Pangu model has been heavily optimized for Ascend computing power, achieving a 30% improvement in training efficiency for native Ascend training. Data from AtomGit shows that as of today, the local offline model openPangu-2.0-Pro has been downloaded 8,834 times on the hosting platform, while openPangu-2.0-Flash has reached 22,068 downloads.
The Pangu Large Model is a crucial strategic branch for Huawei in building the "silicon-based black soil for an intelligent world." Previously, on September 17, at Huawei Connect 2026, Wang Tao, Huawei's Rotating Chairman and Deputy Chairman, announced that the core of Huawei's AI strategy is computing power, adhering to hardware monetization; building an open-source and open computing power ecosystem to support native training of mainstream large models and actively embracing "a thousand models in diverse forms"; and simultaneously applying the Pangu Large Model to the intelligence of Huawei's products to enhance its own product competitiveness.
Today, Huawei further released the complete training stack for openPangu-2.0, marking a critical step in concluding the open-sourcing of the openPangu-2.0 project.