alexsjones/llmfit v1.1.17 is a Rust command-line tool that finds large language models you can run on your computer's specs. With a single command, it surveys hundreds of models and providers so you can pick candidates to run locally. The latest release is v1.1.17.
The tool checks CPU and system memory, GPU, VRAM, and accelerator configuration to recommend models suited to each quantization level. Quantization is a way of reducing a model's representation so it can run with less memory. Because it looks at several components together, users get a starting point for narrowing down candidates on their own equipment. The feature the README particularly emphasizes is the flow for sharing measured benchmarks. You download and run a model, measure tokens per second (tok/s), and then submit the results as a full request (pull request) in the TUI (terminal user interface). Anyone can contribute without an external gh CLI or a separate account, and measurements from the same hardware appear in the next release. So the key is a structure that does not stop at recommendations but collects real measured numbers. It is also meaningful that you can distinguish between estimates and figures obtained from actual runs.
That said, the idea of recommending models that fit your hardware is already common. What distinguishes this tool is the link between recommendations and directly measuring and sharing them. The accuracy of hardware detection and fit calculations is the tool's own judgment, and the provided information contains no comparison results proving recommendation accuracy. The recommendation list itself does not mean measured results. So rather than treating the output as final values, it is better to use it to narrow down candidates and judge real performance by numbers measured on your own hardware.