RigMatch

---

What is RigMatch?

Picking a local LLM is confusing: parameter counts, quantization, VRAM, context windows… RigMatch skips all of that. It benchmarks models on your actual hardware through Ollama or LM Studio, scores each one on speed, answer quality, and hardware fit, and crowns a Top Match with a 0–100 Match Score.

Everything runs on your machine. No account, no telemetry, and no cloud unless you turn on the OpenRouter judge or test one of Ollama's cloud models (see Privacy).

RigMatch is donationware. Every feature is free, and nothing is locked if you don't donate. If it helps you, you can buy me a coffee.

RigMatch is source-available, not open source. The code is public so you can check what the app does, but the license doesn't let you copy or redistribute it (see License).

• 🖥️ Reads your real rig. GPU, VRAM, RAM and disk decide which contestants even qualify, and the Models list in Advanced Mode gives each model an estimated Fit grade from A to F before you download it
• 💛 Speed Dating benchmarks. Every model answers the same questions, live on stage
• 🏆 Match Score (0–100) — speed, quality and fit, scored on your PC
• 💬 Chat with the winner in RigMatch Chat, or skip the show and start chatting
• 🖼️ Pictures, video and sound through ComfyUI, if you have it
• 🎨 5 color themes and illustrated contestant avatars for every model family

Simple Mode — local AI for everyone

A five-step guided wizard for people who have never heard of VRAM and just want "the best AI for my PC." A game-show host walks you through the whole thing.

1 · Setup — check your computer

RigMatch finds Ollama or LM Studio and reads your graphics card, memory and free disk space. No jargon, just "Your computer is ready."

Just want to chat? The card under Choose your models skips the show. It offers a good all-round model for your PC: if you already have it, the chat opens straight away; if not, it tells you how big the download is and asks before getting it. Pick has the same way out, a Skip the show link.

2 · Pick — meet the contestants

Choose what you want a model for (Chat, Writing, Code, Reads pictures, Listens, Makes pictures, Makes video, Makes sound) and pick up to five contestants that fit your PC. Each card says whether the model is already on your PC or how big its download is. One card per model; RigMatch picks the right size for your hardware. Not sure? Choose for me picks a lineup for you.

3 · Download — the contestants arrive

RigMatch asks before it downloads anything, then shows live progress while your lineup installs. Stop anytime; downloads resume if you close the app. If every model in your lineup is already on your PC, this step is skipped.

4 · Show — Speed Dating, live on stage

One sheet before the start says which questions run, who marks them and how hard your computer will work. Then every contestant answers the same questions on the game-show stage, one at a time, with live scores. One of the questions asks each model to draw a pelican riding a bicycle, as a drawing written in code (SVG), an idea from Simon Willison's pelican benchmark. Stop ends the whole show.

5 · Winner — your Top Match

The reveal: your Top Match under the lights with its grade, Match Score and the four measurements behind it. Rank by moves the crown between Accuracy first, Balanced and Speed first. Then two doors out (start chatting right away, or open the control room, which is Advanced Mode) and how the whole lineup finished.

Chatting with your models

RigMatch Chat is the chat window that comes with RigMatch. Open it from Chat in the top bar, from the Winner screen, or from the Just want to chat? card. It lists your Ollama models with their Match Scores, and it remembers only what you ask it to keep, in a "What it remembers" panel you can read and add to. Ask a second model too sends each message to two models and shows their answers side by side; "Keep this one" goes back to a normal chat with the one you preferred. If the two don't both fit in graphics memory, Chat warns that they may take turns. LM Studio models chat in a panel inside RigMatch instead, because RigMatch Chat lists only Ollama's models.

Pictures, video and sound

Makes pictures, Makes video and Makes sound run in ComfyUI, a separate free program that RigMatch doesn't install or bundle. RigMatch looks for it by itself on port 8188 (the portable and manual installs) and 8000 (ComfyUI Desktop); if yours runs somewhere else, set its address in Settings. Pick shows the makers that fit your graphics card, and Simple Mode can download picture makers such as Z-Image Turbo and FLUX.2 [klein], after telling you how much each one fetches. Without ComfyUI, Simple Mode explains how to get it.

Advanced Mode

The full control room, with its tabs in the top bar. Your Top Match carries over.

• Models: every model RigMatch knows, each with an estimated Fit grade for your PC (A to F, with the reasons a click away; a question mark when its size isn't known), filters for what it can do, plus downloads
• What's New: models that just appeared in Ollama's catalog
• Comparison: run a show with the models, questions and judge you choose, and add skill tests such as Build an app or the Code challenge. Turn on Blind show and the contestants become Contestant 1, 2 and 3, with names, sizes and scores hidden; at the end you pick your favorite from their answers side by side, then RigMatch shows who each one was and which was the Match winner
• Labs: picture, video and sound makers given the same prompt, and listening tests
• Results: your Top Match and its measurements, the ranking, and every test you've run, with a fader between accuracy and speed. Your taste ranks models by your blind picks alone, kept apart from the judge's scores
• My PC: your hardware, Setup Doctor and logs, and Speed score (RigScore): the same test on any computer, giving one number you can compare. Higher is faster. An RTX 4070 scores about 1,000 in the app; the same PC can read 15 to 20 percent higher or lower from run to run, and the command-line test usually reads a little higher. A heavy tier drops a lot when another program is holding the graphics card
• Settings (the gear): one section at a time, picked from a list

Getting started

1. Install Ollama and make sure it's running. Already use LM Studio? RigMatch tests the models you have there too, while LM Studio's local server is running. Downloads go through Ollama.
2. Download RigMatch for your platform:

Platform	Installer
Windows	.exe installer or .zip portable · Releases
macOS (Apple Silicon)	.dmg for M-series Macs · Releases
macOS (Intel)	.dmg for Intel Macs · Releases
Linux x64	.AppImage or .deb for x64 Debian/Ubuntu · Releases
Linux ARM64 / Jetson	.AppImage or .deb for ARM64/aarch64 · Releases

3. Launch it — Simple Mode will check your computer and take it from there

RigMatch's Windows builds are not code-signed, because a signing certificate costs money every year and RigMatch is free. So the first time you run the installer (or RigMatch.exe from the .zip), Windows SmartScreen may say "Windows protected your PC".

1. Click More info. It shows the file name and Publisher: Unknown publisher.
2. Click Run anyway.

To check that your download is the one GitHub built, compare its SHA-256 with the line for that file in SHA256SUMS.txt on the same release page (upper or lower case doesn't matter):

Get-FileHash .\RigMatch-*-win-x64.exe -Algorithm SHA256

Or, with the GitHub CLI, check the build attestation:

gh attestation verify RigMatch-<version>-win-x64.exe --repo DaveEuson/RigMatch

RigMatch macOS downloads are unsigned beta builds distributed outside the App Store. On first launch, macOS may say the developer cannot be verified.

1. Download the correct .dmg: mac-arm64 for Apple Silicon, mac-x64 for Intel.
2. Open the .dmg and drag RigMatch to Applications.
3. First launch only: right-click RigMatch.app, choose Open, then Open again.
4. If macOS still blocks it, open System Settings → Privacy & Security → Security and choose Open Anyway.

If macOS says the app is "damaged and can't be opened", and nothing appears in Privacy & Security to let you open it anyway, run this once:

xattr -dr com.apple.quarantine "/Applications/RigMatch.app" && codesign --force --deep --sign - "/Applications/RigMatch.app"

Both halves matter on Apple Silicon. Clearing the quarantine flag alone is not enough: an arm64 app will not launch unless its code signature validates, and that is the part that was broken. The second command re-signs it locally.

This affects builds up to and including 0.4.2. Later releases sign themselves during the build, so "damaged" should not appear at all. You may still see the ordinary unidentified-developer prompt covered in the steps above.

Pick the file for your processor: amd64 / x86_64 for most PCs, arm64 for NVIDIA Jetson and other ARM64 boards.

.deb (Debian, Ubuntu, JetPack): install it through apt, which also installs everything RigMatch Chat needs:

sudo apt update
sudo apt install ./RigMatch-*-linux-amd64.deb

On a Jetson or another ARM64 board, the file ends in linux-arm64.deb instead.

Double-clicking the .deb can open your software store instead. On Pop!_OS 24.04 the COSMIC Store shows an Install button that does nothing, so use the command above.

AppImage: make it executable, then open it:

chmod +x RigMatch-*.AppImage
./RigMatch-*.AppImage

An AppImage installs nothing, so RigMatch Chat uses your system's WebKitGTK 4.1. If it is missing, RigMatch says so when you open Chat; on Debian and Ubuntu:

sudo apt install libwebkit2gtk-4.1-0

AppImages up to 0.9.1 also need libfuse2 to open, which Ubuntu 22.04 and newer do not install: sudo apt install libfuse2 on 22.04, libfuse2t64 on 24.04. Later AppImages do not need it.

Build from source

git clone https://github.com/DaveEuson/RigMatch.git
cd RigMatch
npm install
npm run dev        # development
npm run build      # production build

Command-line fit check

The Fit grade from the Models list, in a terminal: which models will run well on this computer, best first, with the command that gets each one. It grades the chat models in a snapshot of Ollama's library bundled with each release (this one from October 2026), reads your graphics card the way the app does, and asks a local Ollama, if one is running, which models you already have. Nothing is downloaded and nothing leaves your computer.

npx rigmatch-fit

It needs Node 20 or newer. To run it from a clone of this repo instead: npm install, npm run build:cli, then node cli/dist/rigmatch-fit.mjs.

Grade  Model             Size    Speed      Fit         Command
A      qwen3.5:9b *      6.1 GB  ~49 tok/s  Sweet spot  ollama run qwen3.5:9b
A      qwen3-vl:8b       5.7 GB  ~53 tok/s  Sweet spot  ollama pull qwen3-vl:8b
A      mistral-nemo:12b  6.6 GB  ~46 tok/s  Sweet spot  ollama pull mistral-nemo:12b
A      ministral-3:8b    5.6 GB  ~54 tok/s  Sweet spot  ollama pull ministral-3:8b
A      qwen3:8b *        4.9 GB  ~61 tok/s  Sweet spot  ollama run qwen3:8b
...
The best 20 of 202 models that run here.

Models marked * are already in your Ollama, so their command is ollama run.

Option	Does
--all	List every model, including coding and other specialist models, and the ones that will not fit
--installed	List only the models Ollama already has, graded
--json	Print every model as JSON, for scripts (with --installed, only those)
--provider lmstudio	Print LM Studio's lms get commands instead of ollama pull
--vram 16	Grade for a card with 16 GB instead of this one
--measure	Measure the real speed with the Speed score test on the tier models you have (takes minutes)
--measure <model>	Measure one installed model's real speed, beside its estimate
--endurance <model>	Find where one installed model slows down or loses its place as the context grows, at 2K, 8K and 16K (a few minutes)

--measure is the speed test from My PC › Speed score, run from a terminal through your own Ollama: Light (llama3.2:3b), Standard (qwen3:8b), Heavy (qwen3:14b) and, on a 24 GB card, Extreme (qwen3:30b-a3b). It prints each tier's writing and reading speed, points against the reference PC (1,000 is an RTX 4070 12 GB) and what the speed feels like, with the Standard tier as the headline. It never downloads a model (a missing one gets its ollama pull command), refuses to start rather than share the card with another program's model, warns when the graphics card is already busy (NVIDIA, any card on Windows, AMD on Linux, Apple Silicon; a card it cannot read is reported as not checked), and unloads what it loaded. --json with it prints the measures and each one's stamp.

Tier      Model          Writing    Reading       Points  Feels
Light     llama3.2:3b    166 tok/s  10,085 tok/s  1,182   Excellent
Standard  qwen3:8b       79 tok/s   4,358 tok/s   1,177   Excellent
Heavy     qwen3:14b      46 tok/s   2,393 tok/s   1,180   Fast
Extreme   qwen3:30b-a3b  Not run: needs 24 GB of graphics memory, and this PC has 12 GB.

--endurance <model> does the same for the context. The speed test runs at a short context, so this fills 2K, 8K and 16K with a prompt that hides a passkey and says where that model slows down, moves part of itself to system memory, or loses the passkey, on your card. It takes the same care as --measure, and it is also in the app, under My PC › Speed score.

A is the best pick for this computer and F will not fit. The grade weighs answer quality (by size and model family), estimated speed, and the room left in graphics memory. It is a different scale from the Match Score, which RigMatch measures by actually running the model. The bundled file runs on Node 20 or later on Windows, macOS and Linux; its only dependency is systeminformation.

How the Match Score works

Each model runs the same question set on your hardware. The score blends:

Ingredient	What it measures
⚡ Speed	Tokens/sec and first-token latency on your rig
🎯 Quality	Whether answers hold up: valid JSON, followed instructions, no invented facts. Fixed rules mark most answers. A judge model marks the open-ended ones: one of your local models, picked for you unless you choose, or an OpenRouter model if you turn that on. A quick check uses rules only
🧩 Fit	How comfortably the model sits in your VRAM/RAM

The result is a 0–100 Match Score and a letter grade, measured on your machine rather than a leaderboard's H100. Speed and fit are measured directly. Quality is the softest of the three, which is why the app shows the breakdown rather than only the total.

Privacy

Models run on your hardware through Ollama or LM Studio. RigMatch has no account and no telemetry, and nothing you type, test or score leaves your computer unless you switch one of these on:

• The OpenRouter judge. If you choose it and add your own OpenRouter key, the questions it marks and your models' answers to them go to OpenRouter, and so do the apps and code it reviews in Labs.
• A cloud model. Ollama's cloud models run on Ollama's servers, so a test of one sends its questions there.

The sheet before every test says what leaves your computer, if anything.

License

Source-available, not open source — see LICENSE.

The code is published so you can read it and check what the app does, which is the point of a privacy claim you can't otherwise verify. It does not grant rights to copy, modify, or redistribute it. Using the released app is fine.

---