llmfit setup: find models that fit your hardware

This tutorial introduces llmfit, a command-line tool for choosing a local LLM based on a computer's hardware. The presenter installs it through Homebrew on a Mac, then shows recommendations for an M4 Pro. The results include model parameters, scores, estimated tokens per second, memory use and fit ratings.

The walkthrough covers searching by model name, filtering by provider or use case, and exporting JSON for scripts or agents. A model-specific info command shows the estimate basis and fit analysis. When a search matches several models, the presenter supplies a more specific name. The benchmark section explains how to measure throughput and time to first token for models already running through Ollama or llama.cpp; it does not present measured benchmark results.

According to the presenter, llmfit detects RAM, CPU and GPU resources and scores models on quality, speed, memory fit and context. It checks quantization options from Q8 down to Q2 to select the highest quality that fits available memory, then retries at half context if necessary. These estimates help viewers decide what to download before they run models locally.

The presenter also reports that the model catalog updates weekly. A newly released model may therefore be absent, and an outdated llmfit installation can also explain missing entries.