The first hour with Ollama usually goes: pull a model that sounds good, wait, watch it crawl at two tokens a second or refuse to load at all, pull a smaller one, repeat. llmfit skips the loop. It inspects the machine — CPU, RAM, discrete or unified GPU memory, CUDA/Metal/ROCm — and scores hundreds of catalogued models on whether they fit, how fast they’d run, and how good they are, then picks the best quantization that fits rather than assuming one.

Rust, MIT, single binary.

Install 

  brew install llmfit                                   # macOS / Linux
curl -fsSL https://llmfit.axjns.dev/install.sh | sh   # anywhere
uv tool install -U llmfit                             # if you live in Python
docker run -it --rm ghcr.io/alexsjones/llmfit --tui   # no install

Use 

  llmfit                       # TUI: browse, filter with /, sort, read the fit score
llmfit recommend             # hardware summary + the short list, as text
llmfit recommend --json      # same, for scripts
llmfit fit                   # every model in the catalog, ranked by fit
llmfit info "qwen2.5-coder:14b"   # why this one does or doesn't fit
llmfit doctor                # what it detected — check this first on odd hardware
llmfit bench                 # measure real tokens/s against a running Ollama/llama.cpp

The four scores are fit (does it load in memory), speed (a bandwidth model calibrated on community measurements, then corrected by bench if you run it), quality, and context. The quantization ladder walks Q8_0 → Q6 → Q5 → Q4 → Q2 and stops at the best one that fits, so a 14B model on a 16 GB Mac comes back as “yes, at Q4_K_M, ~18 tok/s” instead of “no.”

It knows Ollama, llama.cpp, MLX, LM Studio, and Docker Model Runner as providers, and GGUF/AWQ/GPTQ/EXL2 as formats, so the recommendation is something you can paste into ollama pull rather than a Hugging Face name to go translate.

The honest ceiling

It’s a catalog plus a bandwidth model. New models take a release to show up, and the speed number is an estimate until bench replaces it with a measurement. Treat the output as a shortlist of three, not a verdict — then pull the top one and run bench.

Next

Run the model it picked → Ollama

Edit this page

Last updated 13 Sep 2026, 00:00 UTC. history