# Turbofit

An adaptive local-inference provider for the Hermes Agent chat bot that inventories a machine's compute and memory, recommends the best model it can safely run, and exposes it through one OpenAI-compatible endpoint.

Turbofit is a Hermes Agent model provider that scans a machine's available compute, whether dedicated VRAM, unified or integrated memory, or RAM only, and launches a native inference backend running the strongest model that setup can sustain, all reachable through one OpenAI-compatible endpoint addressed as a single named provider. Its TurboFit Check step applies an automatic hardware lane or a manually selected one, and its evidence-backed model list ranks five models suitable for running Hermes Agent locally by capability, alongside a broader model zoo of additional options including a music-generation model.

A fit rule governs which model a given tier can run: model size plus its key-value cache at long context must stay within that tier's memory, and a near-lossless quantized configuration is used as the default for GPU tiers.

It is a Python project, and a project wiki documents the full model zoo, benchmarks and hardware tiers in more detail.

Website: https://github.com/SouthpawIN/turbofit