Huawei open-sources Pangu 2.0 full training stack, from pretraining to RL code

智东西 · wechat · 2026-09-28

Huawei has released the full training code for openPangu-2.0, its open-source Pangu model, covering pretraining, SFT, and post-training reinforcement learning — optimized for Ascend clusters and marking the final step of the open-source project.

The openPanguTrainingFramework supports LLMs, vision-language, and multimodal models with distributed parallelism, long-context training, mixed precision, and custom operator optimization. The RL framework openPangu-2.0-RL uses VERL as its orchestration kernel with Ascend-affine patches, enabling Actor/Rollout/Reward co-training with GSPO/GRPO.

Huawei previously open-sourced the 92B-param Flash model in June and the 505B-param Pro in July, with Ascend-native training efficiency up 30%. Downloads on AtomGit: 8.8K for Pro, 22K for Flash. The release aligns with Huawei's compute-centric, hardware-monetization AI strategy.

Original post →

More from Companies & People

Companies & People channel →