MMF — Machine Model Fit
Find the best AI model for your machine.
MMF is a small, fast desktop app that inspects your hardware and tells you
which local LLMs you can realistically run — at what quantization, context
length, and expected speed — with one-click install into Ollama.
modelfit.docgraph.app · A DocGraph product.
Why
"Can my machine run Qwen3 32B?" is answered today by Reddit threads and VRAM
calculators. Most tools stop at does it fit. MMF answers the three
questions that actually matter:
- Can the model run? — weights + KV cache + overhead vs. your usable memory
- How fast will it run? — estimated tokens/sec from memory bandwidth;
"fits" and "usable" are different things
- What's the best model for this machine and task? — quality × speed,
per objective (overall / quality / speed / coding)
The answer is always three picks: BEST, SAFE, and FAST — with every
excluded model explaining why ("needs ~26 GB, your usable memory is 21 GB").
Highlights
- Zero prerequisites — no runtime, no Python, no Node needed. Install the
~10 MB app and get recommendations immediately.
- Real hardware detection — Apple Silicon unified memory, GPU cores,
VRAM, acceleration backends (Metal/CUDA/ROCm). Detected specs are editable.
- Correct math where naive calculators fail — MoE models budget memory by
total parameters but speed by active parameters; KV cache is computed from
your requested context length, not a flat guess.
- Measured, not just estimated — one click benchmarks a small model you
already have (or a 2 GB fallback), derives your machine's real effective
memory bandwidth, and re-extrapolates every speed estimate from it. Time to
first token is reported only once that benchmark has run.
- Every number says how well it's known — measured here / calibrated /
estimated / unverified, each hoverable for the inputs behind it
("7.1 GB weights + 2.8 GB KV cache at 8k context + 1.5 GB runtime overhead").
A guess and a measurement never look alike.
- One-click install — pulls the recommended model into
Ollama with live progress.
- Data-driven model registry — models live in a remotely-updated
registry.json, not in code. New models arrive without an app update. It has
two tiers: curated entries carry a hand-written quality score and are the
only ones ever offered as a pick; discovered entries are found
automatically and listed with real memory and speed, but never recommended —
because "best for you" is a claim only a curated score can carry.
Status
Early development. Working today on macOS (Apple Silicon):
Architecture
apps/desktop/ Tauri 2 + React + TypeScript + Tailwind
crates/
hardware/ native detection (sysinfo + platform APIs)
registry/ model-facts schema + bundled snapshot
recommendation/ pure-math engine: constraints → weighted score
runtime-adapters/ RuntimeAdapter trait + Ollama impl (llama.cpp: v2)
share/ benchmark-share payload + Copy diagnostics report
registry/ bundled registry.json snapshot
registry-pipeline/ CI-only Python: models.yaml (curated) +
discover.py → discovered.yaml → registry.json
The engine is an embeddable Rust library — the desktop app, a future CLI, and
other DocGraph products are thin frontends over the same crates. A runtime
(Ollama) is only the executor of a recommendation, never a dependency of
producing one.
See REQUIREMENTS.md for the full product spec and design
decisions.
Building from source
Prerequisites: Rust (stable), Node 20+, and the
Tauri prerequisites for your OS.
# run tests
cargo test
# run the app in dev mode
cd apps/desktop
npm install
cargo tauri dev
# build a release bundle
cargo tauri build
License
MIT © DocGraph
