Models specialised in programming. Code differs from prose in useful ways: it has strict syntax, it can be executed, and correctness can often be checked automatically. That last property is why progress here has been faster than in most other areas.
Technical view
Trained on large code corpora with objectives including fill-in-the-middle, which supports completion with surrounding context rather than only continuation. Evaluated by functional correctness against test suites rather than text similarity, and increasingly on repository-level tasks requiring changes across multiple files. Repository context, dependency awareness and build environment access matter more to practical performance than raw benchmark scores on isolated functions.
Related
Part of AI Terminology Index, section: Modalities and Applications.