Getting started

Mila is a source distribution. C++23 modules cannot ship as prebuilt binaries — the consumer compiles the module graph in its own toolchain — so getting started means building it.

1. Check the toolchain

Earlier Visual Studio 2026 builds have a C++23 module regression; 18.6.2 fixed it. GCC 15.2 and earlier cannot compile the modules at all.

2. Build and test

git clone https://github.com/toddthomson/mila.git
cd mila
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DMILA_ENABLE_TESTING=ON
cmake --build build
ctest --test-dir build

Tests are opt-in — drop -DMILA_ENABLE_TESTING=ON for a library-only build. On Windows you can instead open the folder in Visual Studio, which detects CMakeLists.txt and the bundled presets automatically.

There is also a dev container mirroring the CI toolchain, if you would rather not install CUDA and a compiler locally.

3. Get model weights

Weights are not included. A converter turns HuggingFace checkpoints into Mila’s binary format, and always writes BF16 — quantization to FP8 or FP4 happens at load time, so there is no separate quantized checkpoint to manage.

GPT-2 is the easiest first target: it is ungated, small, and needs no HuggingFace account. Llama and Gemma are gated and require auth.

4. Run something

The Chat CLI is the quickest way to see it work — an instruction-following harness with model hot-switching and tool calling, defaulting to Gemma 4 12B at FP4.

5. Or consume it from your own project

CMake FetchContent is the supported path. Your project compiles Mila’s modules in your toolchain, then links against them like any other target.

The full guide — every step above with exact commands, converter setup, the gated-model auth flow, FetchContent details, and how to open a first pull request — is getting-started.md in the repository. It is the authoritative version; this page is the short road in.