PanGu-Coder

PanGu-Coder is Huawei Noah's Ark Lab's entry in the first wave of code LLMs, built on the PanGu-Alpha decoder architecture. Its distinguishing idea was a two-stage training strategy: causal language modeling on raw code first, then a mix of causal and masked objectives focused on function-level generation from natural-language descriptions, trained on loosely curated docstring-function pairs. A further fine-tuned variant, PanGu-Coder-FT, targeted competitive programming and CI-tested code.

The paper reported the 2.6B model performing on par with or better than similarly sized models like Codex, despite a smaller context window and less training data. The weights were never publicly released, so it survives as a research reference rather than a usable model. It is still worth cataloging as the forerunner of Huawei's later, bigger PanGu-Coder2 line.

Provider-reported: equivalent or better than similarly sized Codex-class models on the authors' function-level synthesis eval; no public HumanEval figures