LMLang

A programming language designed for AI authors.

Released Programming language

LMLang is token-minimal, model-agnostic, content-addressed and contract-verified. Humans read the contracts and capability headers; models read and write everything else.

What it does

The premise is that a language written for a model to author should optimise for different things than one written for a person to type. So: minimal token cost per definition, contracts the compiler can check rather than comments it ignores, content addressing so identical work is deduplicated, and capability headers that make what a function can reach explicit.

The CLI parses, typechecks, capability-checks and executes. the verify command runs a ten-thousand-case contract gauntlet, or proves the contract outright with an SMT solver rather than sampling it. the resolve command fills declared intent holes from the store or by verified synthesis. Builds target an exact-closure artifact or WebAssembly — a Fibonacci module comes out at 91 bytes.

The reference implementation is Rust: 18 crates, 867 tests, and three complete applications written in LMLang itself.

Who it is for

Language and compiler researchers, and anyone building systems where a model writes code that has to be checkable.

What it costs

Open source.

How we price everything else →

Terminology

Words this page uses in a specific sense.

Contract verification
A property attached to a definition that the toolchain checks — by generating ten thousand cases, or by proving it with an SMT solver.
Content-addressed
Every definition is identified by the hash of its contents, so identical code is stored once and any build is reproducible.
Capability header
A declaration of what a definition is allowed to reach. Anything undeclared is refused at check time, not at runtime.
Intent hole
A gap left in a program, written as a question, that the resolver fills from the store or by synthesis — and then verifies against the surrounding contracts.

Facts

FieldValue
NameLMLang
StatusReleased
CategoryProgramming language
PlatformsWindows, Linux, macOS, WebAssembly
PricingOpen source.
SourceClosed source
PublisherCognitive Industries
Published fromBrisbane, Queensland, Australia

Capabilities behind it

Systems and low-level

A from-scratch Rust operating system with capability security and deterministic replay, and an intent runtime whose formal proofs run in microseconds.

AI and machine learning

Models trained from scratch, self-improving reasoning engines, evolutionary architecture search, plus the shipped products that lean on them.

Quantum and heterogeneous compute

A provider-agnostic intermediate representation that plans each piece of work onto CPU, GPU or quantum hardware with cost estimation across real QPU providers.

Related

ChazzAI

AI roleplay and living worlds.

Unslop

Find and fix known bugs. No model in the loop.

HumanOR

Humanize AI text and flag writing that reads machine-made.

Last reviewed · Site changelog