InCoder

InCoder, developed by Meta AI Research, represents a significant leap in the evolution of code-centric large language models. Unlike traditional autoregressive models that generate code purely from left-to-right, InCoder is architected to perform bidirectional code infilling. By training on a specialized causal masking objective, the model can predict missing segments of code based on both preceding and succeeding context. This makes it exceptionally powerful for tasks like refactoring, bug fixing, and type inference where the surrounding logic is already defined. Built in 1.3B and 6.7B parameter variants, it was trained on a massive repository of 28+ programming languages from public GitHub repositories. In the 2026 market landscape, InCoder serves as a foundational architecture for specialized IDE plugins and automated refactoring agents that require more nuance than standard completion. While newer models have emerged, InCoder’s specific focus on the infilling objective remains a benchmark for technical tasks requiring middle-of-file synthesis without losing contextual coherence.