InCoder

InCoder, published by Meta AI in April 2022 (arXiv:2204.05999, ICLR 2023), introduced a simple but influential idea: train a decoder-only transformer with a causal-masking objective so that left-to-right generation and code infilling are the same capability. By moving masked spans to the end of the sequence during training, the model learned to fill in missing code conditioned on both left and right context — the interaction pattern behind IDE autocomplete, bug fixing, and code editing.

Released in 1.3B and 6.7B sizes, it was trained on public code from GitHub and GitLab restricted to permissive, non-copyleft licenses plus StackOverflow, spanning 28 languages with a Python/JavaScript emphasis. It became the reference infilling model of its era and the baseline SantaCoder was measured against. Weights are publicly available; verify the current license terms on the Hugging Face model card before commercial use, as Meta's research releases carry specific conditions.