Can it run?
Weights plus KV cache plus runtime overhead, against the memory your machine can actually give a model — 70 % of a unified pool, or your dedicated VRAM.
Local AI advisor
Find the best AI model for your machine.
“Can my laptop run a 32B model?” is answered today by forum threads and spreadsheet calculators. DocGraph ModelFit reads your actual hardware, does the memory maths properly, and can measure your machine's real speed — then tells you what to run, and installs it.
Free · MIT-licensed · macOS today, Windows & Linux in progress
Why another calculator
Most tools stop at “does it fit”.
That's the easy third of the question.
Will it be usable?
Fitting and being pleasant to use are different things. Generation speed is estimated from memory bandwidth — the number that actually predicts tokens per second — so a model that would crawl is called out, not recommended.
Which one is best?
Quality traded against speed for what you're doing — chat, reasoning or coding. You get three answers: Best, Safe and Fast, and every model that didn't make it says why.
How it works
Four steps, and only one of them needs the internet.
-
1 Read the machine
CPU, cores, memory, GPU and its cores, free disk, and which acceleration backends exist — Metal, CUDA, ROCm. Apple Silicon's unified memory is treated as what it is: one pool the GPU can see. Anything detection gets wrong, you can correct by hand.
-
2 Load the model registry
A small database of model facts — parameter counts, per-quantization file sizes, KV-cache cost, capabilities — built automatically from GGUF metadata and refreshed without updating the app. A copy ships inside the app, so a first run works offline.
-
3 Do the maths that others skip
KV cache is computed from the context length you ask for, because the same model fits at 8k and doesn't at 64k. Mixture-of-experts models are charged memory for all their parameters but speed for only the active ones — get that backwards and a 30B-A3B looks slower than a 14B, which it very much is not.
-
4 Measure, don't guess
One click benchmarks a small model you already have (or, if you have none, downloads a small one of about 2 GB first), times real generation, and derives your machine's true effective memory bandwidth. Every estimate on screen is then re-derived from that measurement. On the M4 Pro above, measuring found 221 GB/s of real throughput, and the top pick, Gemma 4 12B, is calibrated to about 31 tok/s from it.
What it is
Nothing to install first
No runtime, no Python, no Node. The app detects and recommends on its own. Ollama is only needed if you want it to measure or install a model for you — and if it's missing, the app says so and carries on.
One outbound request
Fetching the model registry. That's the only thing DocGraph ModelFit sends anywhere. Your hardware details never leave the machine — there is no account, no telemetry, no upload.
Small and native
Rust core, native detection, a webview for the interface. Tauri rather than a bundled browser, so the download is measured in megabytes.
Open, and open to correction
MIT-licensed, engine and registry both. Adding a model is one entry in a YAML file — the pipeline works out its sizes from the GGUF headers itself.
Honest status
Early, and working.
DocGraph ModelFit is in active development in the open. There are no packaged downloads yet — the first signed releases land with cross-platform builds. Until then it builds and runs from source in a couple of minutes.
- Hardware detection macOS, Apple Silicon
- Recommendation engine MoE- and context-aware
- Objectives, context control, full model table
- Ollama: status, one-click install, real benchmark
- Auto-updating model registry
- Windows & Linux GPU detection
- Signed installers and auto-update
Build it now
git clone https://github.com/docgraphapp/modelfit
cd modelfit/apps/desktop
npm install
cargo tauri devNeeds Rust, Node 20+ and the Tauri prerequisites. Watch the repository for the first release.