Advanced AI Coaching System Requirements

Macro Trainer features a state-of-the-art Advanced AI Coaching system. To guarantee complete user privacy, the coaching engine runs entirely locally on your device. No conversations, stats, or logs ever leave your machine or get sent to third-party APIs.

While mobile platforms (iOS and Android) support this out-of-the-box using integrated LiteRT runtimes, desktop clients run the model using the llama.cpp inference engine.

Supported Platforms

No platform needs anything installed by hand:

• iOS & Android: No external programs required. The application runs natively on-device.
• macOS: The llama.cpp runtime ships inside Macro Trainer, so there is nothing to install.
• Windows: The llama.cpp runtime ships inside Macro Trainer, so there is nothing to install — turn on Advanced AI Coaching, download the model, and it works.
• Linux: Macro Trainer downloads its own copy of the llama.cpp runtime (a one-time ~15 MB download) when you enable Advanced AI Coaching. This works in all install formats, including the sandboxed Snap and Flatpak packages.

Desktop Setup

macOS Setup

No setup is needed, and nothing is downloaded: the llama.cpp runtime is installed inside Macro Trainer itself, and the app uses it automatically. It is built for Apple Silicon and uses Metal, so the GPU does the work and replies stream back in a second or two.

If you have your own llama.cpp install: Macro Trainer prefers its own bundled copy and will not use yours. Earlier versions relied on a Homebrew install instead; if you installed llama.cpp only for Macro Trainer, you no longer need it (brew uninstall llama.cpp).

If Macro Trainer tells you the runtime is missing, that means this copy was built without it — reinstalling the latest version from macrotrainer.com will restore it.

Windows Setup

No setup is needed, and nothing is downloaded: the llama.cpp runtime is installed alongside Macro Trainer itself. It runs on the CPU and automatically uses the fastest instruction set your processor supports (SSE4.2 through AVX-512), so a modern processor will be noticeably quicker than an old one.

If Macro Trainer tells you the runtime is missing, that means this installation was built without it — reinstalling the latest version from macrotrainer.com will restore it. There is nothing for you to install by hand.

Linux Setup

No setup is needed: Macro Trainer offers to download the llama.cpp runtime automatically the first time you enable Advanced AI Coaching, and keeps it in the app’s own data directory. The bundled runtime runs on the CPU and automatically uses the fastest instruction set your processor supports (SSE4.2 through AVX-512).

Optional — custom builds: If you installed Macro Trainer from the plain tarball (not Snap or Flatpak) and have your own llama.cpp build on your PATH (for example a Vulkan or CUDA build for GPU acceleration), Macro Trainer will detect and use it, and will skip the runtime download entirely. Sandboxed Snap and Flatpak installs cannot see programs installed on the host system, so they always use the automatically downloaded runtime.

Hardware & System Requirements

• Model File: The application automatically downloads the Gemma-4-2B-IT (Q4_K_M) model in GGUF format.
• Free Disk Space: Minimum of 4.0 GB of free disk space is required to download and save the 3.1 GB model file.
• System Memory (RAM): Minimum of 8 GB, since the model is held in memory while your coach is answering.

Where the work happens

• macOS: On the GPU. The bundled runtime is built for Apple Silicon with Metal, and offloads the whole model — this is the fastest desktop configuration.
• Windows and Linux: On the CPU. The bundled and downloaded runtimes are CPU builds, so reply speed tracks your processor rather than your graphics card. Both pick the fastest instruction set your CPU supports automatically.
• Linux, with your own build: If you run a non-sandboxed install and have your own GPU-enabled llama.cpp (Vulkan or CUDA) on your PATH, Macro Trainer uses it instead — see the Linux section above.