PolyCoder

PolyCoder was built by CMU researchers to fill the gap of a large open-source model trained exclusively on multilingual code. The 2.7-billion-parameter model uses a GPT-2 architecture trained on 249 gigabytes of code across 12 programming languages, on a single lab machine.

Published at the ICLR 2022 Deep Learning for Code workshop, it outperformed all evaluated models including Codex on C-language completion, demonstrating that open models could rival closed frontier code models in specific languages. It is fully open-source under the MIT license.