llmfit

llmfit is a Rust-based terminal app that profiles your machine (CPU cores, system RAM, GPU model, VRAM, available backend) and then ranks hundreds of open-weight LLMs by how they will actually perform on that hardware. It scores each model across quality, speed, fit, and context dimensions, picks the best quantization, estimates tokens-per-second, and flags whether the model is a GPU fit, a CPU offload, an MoE expert-switch case, or a no-go. The default interface is a Vim-keyed TUI with filters by provider, use case, capability, license, and runtime; a classic CLI mode emits JSON for piping into jq.

Where most "what model should I run" guides are blog posts that age out in a month, llmfit is data-driven and live: it integrates with Ollama, llama.cpp, MLX, Docker Model Runner, and LM Studio to mark which models you already have installed, has a Plan mode that inverts the question to "what hardware would I need for this config?", a Hardware Simulation mode to test specs you don't own, and a Community Leaderboard view powered by localmaxxing.com that surfaces real measured tok/s, TTFT, and VRAM from other users on the same GPU. It's MIT-licensed, free, and installable via Homebrew, Scoop, MacPorts, uv/pip, Docker, or cargo.

The project is actively developed by Alex Jones with a tunable scoring panel (efficiency, GPU/CPU/MoE offload factors) so power users can correct for known overestimation cases, plus a download manager that handles model fetching and deletion against the configured runtime. It pairs with sister tools llmserve and llama-panel for actually serving the chosen model.