Llmfit

Why it's ranked here

The estimates show their working. Each speed figure carries a confidence label saying whether it was measured on this machine, measured by others on matching hardware, calibrated, or pure formula, and the inputs behind it are exposed so a number can be checked. The fit bands are explicit (60, 85 and 98 percent pool utilization), mixture-of-experts models are sized by their active parameters rather than their total, and paths that spill off the graphics card can never earn the top verdict. Add an MIT license, a single Rust binary, and installs through Homebrew, Scoop, MacPorts, pip and Docker, and it is an easy tool to trust and to adopt.

What's good

Hardware detection is broad and specific about its limits. It covers NVIDIA, AMD, Intel Arc, Apple Silicon and Ascend accelerators, sums memory across multiple cards, and corrects for AMD laptop chips whose BIOS hides part of the RAM from the operating system. It talks to five local runtimes (Ollama, llama.cpp, MLX, Docker Model Runner, LM Studio), marks what you already have installed, and can pull a model from the terminal interface. Machine-readable JSON output on every command, documented exit codes and an HTTP endpoint make it usable from scripts and agents, not only by hand. A profile flag scores models against hardware you do not own.

Tradeoffs

Most numbers are estimates until someone benchmarks. Speed comes from a bandwidth formula with a fixed 0.55 efficiency factor, and prompt processing time is left blank unless the card's compute throughput is known. The model catalogue is compiled into the binary, so new releases arrive only when you upgrade the tool. The quality score leans partly on a model family's reputation. Detection is thinner off Linux and Apple Silicon: Windows and Intel Macs see NVIDIA cards only through its own utility, and Android phones get no GPU detection at all. Windows release binaries can ship unsigned if the signing job fails.

How to use it well

Run it before choosing a model for a laptop, workstation or home server, and treat the Too Tight and Marginal verdicts as the useful output: they save a doomed download. Pick the use case (coding, reasoning, chat) so the weighting matches your work, and run its benchmark command once a model is serving so your real throughput replaces the estimate. On a machine whose detection misreads memory, override it by hand. It is a sizing tool, not an inference server or an evaluation harness: it will not tell you whether a model answers your questions well, only whether it will run.

Observed

License

MIT

Language

Rust (Cargo workspace, edition 2024)

Interfaces

Terminal UI (default), CLI with JSON and CSV output, web dashboard, HTTP JSON API

Install surface

Homebrew, Scoop, MacPorts, uv/pip, install script, multi-arch container image, cargo build from source

Platform support

macOS (Apple Silicon and Intel), Linux (x86_64 and ARM64), Windows (x86_64)

GPU detection

NVIDIA, AMD, Intel Arc, Apple Silicon, Ascend; no GPU detection on Android/Termux

Runtime providers

Ollama, llama.cpp, MLX, Docker Model Runner, LM Studio

Model catalogue

Embedded in the binary at compile time; updated by upgrading the tool

Speed estimate provenance

Every figure labelled measured locally, measured by community, calibrated, estimated or unsupported