Back to projectsExperimental programming language

ain.

Software for machine authors

Less code.
More capability.

A general-purpose language designed for code written by AI, not optimized for human typing.

Source Compiler releases

Prefix syntaxInferred typesBroad standard libraryRust + experimental LLVM

Language

Only what matters.

example.ainsource

add a b = + a b
square x = * x x

print add square 3 square 4

Output25

Fewer symbols. No call parentheses. No mandatory function keyword.

The compiler handles the structure; the model focuses on expressing the operation. That is the experiment, not a claim that Ain already outperforms other languages.

Read language reference

Design

Deliberately excessive.

Small expressions

Fixed-arity prefix syntax and inferred types reduce the scaffolding a model has to generate.

An enormous vocabulary

An extensive built-in library aims to express complex operations in very little source code.

Real compilation

Rust-backed native compilation today, with a limited experimental AIR and LLVM path.

Token benchmark

Less code. Fewer tokens?

Source-code token counts from the latest published Ain benchmark snapshot, refreshed from GitHub approximately every hour.

Published results

5 matched programming tasks

Tokenizer: o200k_base

Ain

98baseline

Python

126Ain −22.2%

TypeScript

150Ain −34.7%

Rust

192Ain −49.0%

Source tokens for the same tasks in four languages
TASK	Ain	Python	TypeScript	Rust
Arithmetic	16	20	24	31
Factorial	23	28	33	43
Generics	19	25	31	38
Lists	23	33	35	48
Strings	17	20	27	32

Lower is better. Small hand-written examples, counted with the same tokenizer. Not LLM prompt, reasoning, retry, or total generation tokens.

JSON source ↗Refresh workflow ↗

Research, not a finished language.

Ain is pre-alpha. The Rust backend supports more builtins than the experimental LLVM backend. Performance and token claims should be judged using real, repeatable benchmarks.

Documentation