llmfit: find the models your Mac can actually run

llmfit reads the CPU¹ cores, memory, GPU² and VRAM³ of the machine it runs on, then ranks the models and quantizations that fit that configuration. The input is your hardware; the output is a sorted list.

The value is that you do not need to know model names. Before settling on a machine you can ask what it would run.

The verdict splits into four axes: whether it fits in memory, how fast it should be, the model's own quality, and how much context it holds. All four appear on one screen.

On a Mac, install it with Homebrew. Ubuntu and Windows use the same tool, which helps when comparing a Mac against a desktop.

Run it with no arguments. Your detected hardware sits at the top and the scored models below, and Enter opens the detail view for the selected model.

If detection looks wrong, run doctor. It prints exactly what it read, which makes a misread memory size easy to trace.

The fit score reads how much of your available memory the model uses. It rewards the 50 to 80 percent band and marks down both tight and wasteful loads, and because a wrong fit invalidates the rest, read it first.

The speed score is estimated from memory bandwidth⁴. The same model shifts with bandwidth, which is why this axis matters when comparing machines.

Quality and context belong to the model, so they barely move between machines. Read those two when deciding which model to pick.

The speed in the list is an estimate. llmfit exposes the inputs behind it through info, and if those inputs differ from your machine the number moves with them.

Measurement is bench. It attaches to a runtime⁵ you already have running and records generation speed and time to first token⁶, so you measure the setup you actually use.

Seeing both exposes where the estimate drifts. That gap is the point of the tool: buying on the estimate alone can invert the ranking once measured.

It handles GGUF⁷, AWQ, GPTQ and EXL2, and it does not assume a fixed width: it walks from Q8_0 down to Q2_K and picks the highest quality that still fits. If full context does not fit, it retries at half context.

For mixture-of-experts models it calculates with active parameters. Using total parameters makes the model look heavier than it is and hides ones that would run.

The difference shows up on large models: even past 100 GB total, a small active slice can still run on a laptop.

Adding --json to recommend returns output a script can consume, useful for pre-deploy checks or building a table.

brew install AlexsJones/llmfit/llmfit
llmfit doctor
llmfit recommend --json

serve opens an HTTP API⁸ that exposes the system profile and the model list, so a dashboard or pipeline⁹ can read it.

llmfit serve --host 0.0.0.0 --port 8787
curl -s localhost:8787/api/v1/system

bench --share contributes a measurement to the repository. The run is saved on your machine first and only leaves when you send it.

People on identical hardware see it in the next release, where an estimate becomes a check-marked measured number.

The catalog can grow while the estimates improve, because every measurement fills a gap.

llmfit does not run models. Downloading and serving stay with a runtime such as Ollama, llama.cpp or MLX¹⁰, and bench attaches to one already running.

It does not guarantee speed either. Other apps using memory or a different driver change the number, so identical hardware has to mean identical conditions.

Other tools measure by running: they pull models through Ollama and benchmark them, but treat expert models as dense and overstate memory. Which one fits depends on what you are checking.

Use the list as a starting point for a buying decision and confirm with bench before committing.

The last piece llmfit leaves open is how fast the machine actually runs. Its speed comes from a bandwidth calculation, so a different quantization¹¹ file or context length¹² moves the real number away from the estimate.

This site collects prefill¹³, token¹⁴ generation and time to first token per machine with the conditions attached, so one model can be compared across machines under the same conditions. Narrow the candidates with llmfit, then check only those against measured values.

Power is not modelled at all. This site records each machine's peak and idle draw separately, on the maker's figures with the source attached, and the electricity calculator turns hours per day plus your existing household usage into a monthly bill across progressive tiers and seasonal rates.