llmfit
Find out which LLMs your hardware can actually run
About llmfit
llmfit is a terminal tool that answers the question that comes before Ollama, LM Studio or llama.cpp: which open-weight models will actually run on this machine, at what quantisation, and roughly how fast. It inspects CPU cores, system RAM, discrete and integrated GPUs, VRAM and unified memory - NVIDIA CUDA, Apple Silicon, AMD ROCm and Intel OneAPI are all detected - then analyses each model's parameter count, context length and quantisation format (GGUF, AWQ, GPTQ, EXL2) to project its memory footprint and tokens-per-second, and scores it on quality, speed, fit and context. The output is a ranked list of what fits, not a list of everything that exists.
It exists because the gap between "I have a GPU" and "I am running a model that suits it" is where most people give up on local inference, and the usual answer is trial and error against out-of-memory errors. The project is written in Rust, MIT licensed, and has 36,019 GitHub stars and 132 contributors since February 2026, with v1.1.15 released on 2026-09-10. There is no paid tier, no account and no pricing page; the website is a single install line. The README states that the program sends nothing over the network unless you explicitly ask it to, and the release binaries are code-signed through SignPath. The estimates are a model of hardware rather than a measurement, which is why the project added community benchmarks: you can run a real model, measure real tokens per second, and contribute the numbers so everyone's estimates improve.
Install it with brew install llmfit, scoop install llmfit on Windows, cargo install, uv tool install llmfit, a one-line curl script, or a Docker image. Run llmfit with no arguments and the interactive TUI opens: it profiles the machine, lists models across hundreds of entries and providers, and shows for each one whether it fits, at which quantisation, and the projected speed. Classic CLI flags print the same information to standard output or as raw JSON for scripting. A web dashboard and a REST API (/api/v1/system, /api/v1/models) serve the same data to orchestrators or deployment pipelines. It knows the local runtime providers - Ollama, llama.cpp, MLX, Docker Model Runner and LM Studio - so a recommendation maps onto something you can actually start. From the TUI you can open community benchmarks, run live inference benchmarks on your own hardware, and share the results back. Multi-GPU setups, mixture-of-experts architectures and dynamic quantisation selection are handled.
- •Hardware Auto-Detection - CPU, RAM, discrete and integrated GPUs, VRAM and unified memory across CUDA, Apple Silicon, ROCm and OneAPI
- •Model Compatibility Engine - Projects memory footprint and tokens-per-second from parameter count, context length and quantisation
- •TUI, CLI, Web Dashboard and REST API - The same recommendations in whichever form your workflow wants them
- •Runtime Provider Awareness - Maps a recommendation onto Ollama, llama.cpp, MLX, Docker Model Runner or LM Studio
- •Community Benchmarks - Measure real throughput on your machine and contribute it so the estimates improve for everyone
- •Cross-Platform and Signed - macOS (Intel and Apple Silicon), Linux (x86_64 and ARM64), Windows; signed release binaries
Developers setting up local inference for the first time on a laptop or a workstation, teams sizing a model to a specific box before buying more hardware, and anyone wiring local models into a coding agent who wants a defensible answer to "which one". Pair it with the runtimes already in this directory. It is a tool for a moment rather than a workflow: once you know what fits, you rarely run it again. The recommendations are estimates, and an unusual GPU or an exotic quantisation can be off by a meaningful margin - treat the benchmark feature as the ground truth and the estimate as the shortlist. It is a single-maintainer project with several sister tools by the same author, which is worth knowing if you are building it into something long-lived.
Pricing
No pricing page found
Free and MIT licensed. No paid tier, no account and no pricing page; installs via Homebrew, Scoop, MacPorts, cargo, uv, pip, curl or Docker.
Checked 2026-09-15
