At 11:43 on September 28, IT Home reported that Huawei said the pretraining code, the SFT code, and the post-training RL code for openPangu 2.0 were now public. The article gives two addresses: one repository for pretraining and SFT, and one for post-training RL. It describes openPangu as Huawei's open-model brand, aimed at Ascend-native training and inference.

[1][2]
Scratchboard with two black rectangles, and a single white scratch on only the left one.
Of two black cards, only one carries a white scratch. It stands for the training repository whose public note is a single line, beside a card this piece does not fill in., AI-generated illustration, not a news photograph

This piece opened the pretraining and SFT repository. The page title calls it a large-scale foundation-model training framework based on PyTorch and CANN. The project description is one sentence: pretraining and SFT code for the Ascend-native openPangu 2.0 models. The page marks Apache-2.0, shows one commit, and shows three stars. The README does not describe a training procedure, a data source, or a machine count.

[1][2]

The post-training RL address is in the IT Home article. This piece does not copy that page's license, star count, or description, because the repository it checked and listed is the training one. Related links under the IT Home piece also point at a 92-billion-parameter Flash model and at a Pro model with a technical report. Those are other pages, and their parameter counts are not used here. What this piece can say is that Huawei announced both kinds of code, and that the training repository it opened has a one-line public note.

[1][2]

要点

  • Huawei announced public pretraining, SFT, and post-training RL code for Pangu 2.0.
  • The training repository's note is one sentence, marked Apache-2.0, with one commit and three stars.
  • The README does not describe the training steps, the data, or the hardware scale.
  • Parameter counts from related links are left out of this piece.