Docs

What you get

Each hoof model is delivered as a standalone executable — a single file that contains the model weights, tokeniser, inference engine, and a built-in web UI. No Python, no GPU drivers, no internet connection required. Double-click and it runs.

The executable starts a local web server and opens a chat interface in your browser. All processing happens on your machine — nothing is sent to the cloud.

Under the hood

Each executable is powered by the hoof runtime — a lightweight inference engine written in Rust. It handles:

• - Loading and decompressing the model weights at startup
• - Tokenisation (byte-level BPE, compatible with LLaMA, GPT-2, and similar models)
• - Transformer inference with SIMD acceleration (AVX2 + FMA on x86)
• - Quantised execution — 8-bit or 4-bit weights, dequantised on the fly
• - LoRA adapter application — finetuned corrections are baked into the model
• - A local web server with streaming generation and chat history

The runtime is not currently available separately — it's embedded in each executable we deliver. This keeps the models self-contained and ensures consistent behaviour across deployments.

The .hoof format

Internally, each executable contains a .hoof model file — a custom binary format that stores everything needed for inference in a single file:

Web API

Each executable exposes a local REST API alongside the web UI. This means you can integrate hoof models into your own applications:

System requirements