Why Elixir Is the Best Language for LLMs
![]()
Large language models achieve their highest performance on Elixir compared to 19 other languages tested on AutoCodeBench, with a 97.5% upper bound and individual models scoring ~30 points higher on Elixir than their overall average, despite Elixir having roughly 10 times less training data than Python. This superior performance persists across different model architectures, reasoning modes, and even base models without instruction tuning, suggesting the advantage stems from Elixir's language design—particularly its immutability and explicit data flow—which aligns with how language models process code rather than from memorization of training examples.
Was this useful?