v0.9.2-beta — Linux Out Of The Box

What's new in 0.9.2 — Linux Out Of The Box

• On Linux, RigMatch Chat opens on a fresh install. Chat is built on WebKitGTK, and the .deb never asked for it, so on a system without it the Chat button did nothing and said nothing. Installing the .deb now installs it too.
• When Chat can't start, RigMatch says why. Before opening Chat on Linux it checks that every library Chat needs is there. If one is missing, it names it and, on Debian and Ubuntu, gives the command that installs it. The AppImage needs this most, because it can't install anything itself.
• The AppImage opens on Ubuntu 22.04 and newer with nothing extra installed. It used to need libfuse2, which those versions no longer include, so double-clicking it did nothing until you installed that by hand. It now uses the newer AppImage runtime, which doesn't need it.
• Electron 42.10, which fixes four high-severity security issues in the framework RigMatch is built on.

Downloads

• Windows: RigMatch-*-win-x64.exe (installer), or RigMatch-*-win-x64.zip — unpack it anywhere and run RigMatch.exe, which installs nothing and skips the SmartScreen prompt below
• macOS Apple Silicon: RigMatch-*-mac-arm64.dmg
• macOS Intel: RigMatch-*-mac-x64.dmg
• Linux x64: RigMatch-*-linux-x86_64.AppImage or RigMatch-*-linux-amd64.deb
• Linux ARM64 / Jetson: RigMatch-*-linux-arm64.AppImage or RigMatch-*-linux-arm64.deb

Already running an older install? It updates itself to this version automatically — nothing to re-download.

Windows first launch

RigMatch for Windows is an unsigned beta, so Microsoft Defender SmartScreen will almost certainly stop it the first time. That warning means the file has no code-signing certificate yet and has not been downloaded enough times to build a reputation — not that anything is wrong with it.

1. Download RigMatch-*-win-x64.exe.
2. If SmartScreen says "Windows protected your PC", choose More info, then Run anyway.
3. If your browser blocks the download instead, choose Keep and, where asked, Keep anyway.

Prefer not to click through a warning at all? Take the .zip from Downloads above instead.

Either way, verify what you downloaded against SHA256SUMS.txt on this release:

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

A signing certificate is on the roadmap; until it is bought and the reputation builds, this warning is expected on every release.

What you can check in the meantime: every file here is published with GitHub build provenance, a signed record of which workflow built it, from which commit. With the gh CLI:

gh attestation verify RigMatch-*-win-x64.exe --repo DaveEuson/RigMatch

That proves the file came from this repository's release workflow and has not been altered since. It is not code signing, and Windows will still show its warning.

macOS first launch

RigMatch for macOS is currently an unsigned beta distributed outside the App Store. On first launch, macOS may say the developer cannot be verified or that the app was downloaded from the internet.

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

After that first approval, RigMatch opens normally by double-clicking. If macOS says the app is damaged after copying it to Applications, run this Terminal command once:

xattr -cr /Applications/RigMatch.app

Apple reference: https://support.apple.com/guide/mac-help/open-a-mac-app-from-an-unknown-developer-mh40616/mac

What's Changed

• Bump electron from 42.8.1 to 42.10.0 by @dependabot[bot] in #42
• Make the Linux packages open on a stock Ubuntu by @DaveEuson in #43
• Prepare 0.9.2 by @DaveEuson in #46

Full Changelog: v0.9.1-beta...v0.9.2-beta

v0.9.1-beta — Everyone Can Follow The Show

What's new in 0.9.1 — Everyone Can Follow The Show

• The show looks like a show again. Simple Mode's Speed Dating had become a plain dark box: the studio photo behind it was loaded all along, hidden under a solid fill, and the marquee lights from Advanced Mode's stage never made it across. The contestants now answer on a lit stage with a spotlight on whoever is answering, the winner is revealed under the lights, and a show that stops turns them down. None of it takes room from the screen, and the scores sit below the stage, unlit.
• The Compare screen shows its progress bar again. On a 1280 × 800 window the progress bar and every answer score were cut off below the stage, with no way to scroll down to them.
• Picking a lineup looks like a dating show. Each contestant's portrait is larger and framed like a publicity photo, without making the cards any taller. The Pick buttons are pink hearts instead of six gold bars competing with Next, and empty seats in your lineup are hearts. Every step sits on a faint heart wallpaper. On laptop screens the steps across the top show their names again (Setup, Pick, Compare, Winner) instead of numbered circles and padlocks.
• Theme music, off unless you turn it on. Switch on Theme music beside the host, or in Settings → Preferences in Advanced Mode, for a corny 1970s game-show tune while the contestants answer, a drum roll and ta-da for the winner, and a sad trombone when a show stops. RigMatch plays it itself, with no audio file and nothing downloaded, and never over a listening round. It uses about 0.15% of one processor core, the same for every model, far too little to change a score. Advanced Mode's live Speed Dating stage has its own switch.
• Show effects, also off unless you turn them on: contestants walk on one after another, an APPLAUSE sign lights up as each one finishes, curtains part for the winner, and hearts float up when you pick someone. All of it stays still when your computer is set to reduce motion.
• "It's a match" has a proper boat. The victory cruise was three grey boxes and a triangle on a pink-to-blue wash. It is now a white cruiser with portholes, a string of lights and a funnel that puffs hearts, sailing a sunset sea to a harp, strings and a violin instead of three beeps. On an 800-pixel-tall screen the picture no longer covers "Romantic cruise launched" underneath it.
• The show goes on when a model can't finish. One failure used to end the whole show: a crashed model runner, a single dropped connection or one question over two minutes, and every model after it never ran, while whoever had already finished was crowned "out of the 3 you tested". That model now sits out, the rest carry on, the winner is picked from the models that finished, and the ones that couldn't are listed with the reason. This applies to Speed Dating in Advanced Mode too.
• A show that stops before anyone finishes says so. It used to move on to the Winner screen, where the host said "We have a match!" over "Run the show to crown your Top Match." It now stays on the show, says what happened, and offers to run it again. Simple Mode also stopped letting a chat or coding show start with one model, which failed the moment it began; anyone with a single model installed hit that on their first run. Those rounds now ask for two.
• The Winner screen crowns the show you just ran. It showed the best score ever saved on this computer, which could be a model from an earlier session that was not in the lineup at all.
• A model that answers nothing is no longer ranked. One that gave an empty answer to every question still scored about 27 on speed and fit alone, and on "Speed first" could place above a model that answered slowly. It is now listed as one that couldn't finish.
• A judge that stops answering is not waited on for every answer. A hung judge held each answer it was marking for two minutes, while the contestant was shown "Thinking it over…" and the time left climbed from one minute to four. The show now says when an answer is being marked and by which model, gives up on a judge after the first time it fails to answer, and says a question is taking a while instead of stretching the estimate.
• Errors in a show say what happened. "Error: 500 Internal Server Error from http://127.0.0.1:11434/api/generate" now reads as Ollama's model runner stopping, often from running out of memory. A connection lost for one answer is no longer blamed on Ollama not being installed, and pressing Stop is no longer called an error.
• Advanced Mode's header no longer calls a model "Best Tested" before it has been tested. An untested pick says so and offers "Test it", and a model that isn't downloaded yet no longer shows buttons that could not work for it. Macs and AMD graphics cards are no longer told their models run on the CPU, and a Jetson no longer shows a raw nvidia-smi error before every test.
• RigMatch Chat opens on older Linux. Built on Ubuntu 24.04, it needed a newer system library than Ubuntu 22.04, Debian 12, Linux Mint 21, Pop!_OS 22.04 or a Jetson on JetPack 6 provide, so it would not open on them. RigMatch is now built to run on all of them, and every release is checked for it before it ships.
• RigMatch fits small screens. The window had a fixed minimum of 1024 × 640 pixels, so on a smaller screen it ran off the edge and part of the app was out of reach. It now sizes itself to the screen it opens on, and the first-run questions fit an 800 × 600 screen.
• Linux software centers can show what RigMatch is. The AppImage now carries a description and screenshots, which the AppImage catalog and software centers read to list it.
• A download that stops halfway is no longer shown as finished. If the connection to Ollama dropped in the middle of a download, RigMatch said the model was ready and on your PC, and the show started without it. It now says where the download stopped, and starting it again picks up from there.
• A screen reader now hears the show. It used to run for three to ten minutes in silence and then say "Report ready"; everything in between was on screen only. It now says when each model starts answering and what the one before it scored, and it says so if the show stops early. The hardware check says its result too, which it never did, because that result appears further down the page.
• Every button in Simple Mode says what it is. At 1280 pixels wide and narrower, which includes a 1080p laptop at 150% scaling, the five step buttons across the top had no name at all, because the label was hidden in a way that also hid it from screen readers. The model cards had nine buttons that all said "♥ Pick"; each now says whose, as in "Pick Gemma4".
• The Pick, running and winner screens have headings, so a screen reader can jump between models instead of reading every card top to bottom. Each model card is headed by its name, and so is the winner.
• Simple Mode fits a window at 400% zoom. At that size, which is 320 pixels, one unbreakable line on the Pick screen stretched the whole window, and "RigMatch" at the top read "R" while the Next button was cut off entirely. It now wraps, and the footer stacks. 200% already worked.
• Two pieces of text were too faint to meet the contrast standard: "Skip tour" in the tutorial, at 3.5:1, is now 6.7:1, and "Picked · Click to remove" on a chosen model is 5.5:1 or better in all five themes. Explanations that lived only in hover tooltips can now be reached another way: the ComfyUI note appears under the goals once you pick one that needs it, and a screen reader reads each goal's hardware note and each card's country and download count. The dotted terms the host explains are easier to click.
• How this was checked: axe-core against the WCAG 2.1 AA rules, a keyboard walk of every screen, and zoom and text-spacing checks, all in the real app through a full run to the winner. It was not tested with a screen reader in someone's hands. If something reads wrong to you, Report a bug says what happened.

Downloads

• Windows: RigMatch-*-win-x64.exe (installer), or RigMatch-*-win-x64.zip — unpack it anywhere and run RigMatch.exe, which installs nothing and skips the SmartScreen prompt below
• macOS Apple Silicon: RigMatch-*-mac-arm64.dmg
• macOS Intel: RigMatch-*-mac-x64.dmg
• Linux x64: RigMatch-*-linux-x86_64.AppImage or RigMatch-*-linux-amd64.deb
• Linux ARM64 / Jetson: RigMatch-*-linux-arm64.AppImage or RigMatch-*-linux-arm64.deb

Already running an older install? It updates itself to this version automatically — nothing to re-download.

Windows first launch

RigMatch for Windows is an unsigned beta, so Microsoft Defender SmartScreen will almost certainly stop it the first time. That warning means the file has no code-signing certificate yet and has not been downloaded enough times to build a reputation — not that anything is wrong with it.

1. Download RigMatch-*-win-x64.exe.
2. If SmartScreen says "Windows protected your PC", choose More info, then Run anyway.
3. If your browser blocks the download instead, choose Keep and, where asked, Keep anyway.

Prefer not to click through a warning at all? Take the .zip from Downloads above instead.

Either way, verify what you downloaded against SHA256SUMS.txt on this release:

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

A signing certificate is on the roadmap; until it is bought and the reputation builds, this warning is expected on every release.

What you can check in the meantime: every file here is published with GitHub build provenance, a signed record of which workflow built it, from which commit. With the gh CLI:

gh attestation verify RigMatch-*-win-x64.exe --repo DaveEuson/RigMatch

That proves the file came from this repository's release workflow and has not been altered since. It is not code signing, and Windows will still show its warning.

macOS first launch

RigMatch for macOS is c...

v0.9.0-beta — What Matters More?

What's new in 0.9.0 — What Matters More?

• Before every test, RigMatch now asks what matters more to you: getting it right, or getting it fast. "Best" used to mean a fixed blend nobody was asked about, set three menus deep, and a video race was ranked by time alone, which crowns a quick clip of the wrong scene. The Balance fader, accuracy at the top and speed at the bottom, sits beside the results and re-ranks them without running anything again. Each kind of test keeps its own setting, and the three old presets are notches at exactly their old weights, so a Match Score nobody re-weights is the number it always was.
• Three rules hold wherever the fader sits. A result that failed its check never wins, so Speed first crowns the fastest result that passed rather than the fastest result. A result nothing judged cannot win on accuracy it never showed. And when nothing in a race was judged at all, the ranking is speed alone and says so. In a real three-model race, Accuracy first put Wan 2.1 1.3B, which drew everything the prompt asked for in six minutes, above Wan 2.2 5B, which took under three and drew a red barn where the prompt asked for a lighthouse.
• Advanced Mode is organized by what you are testing. It used to show every kind of test at once and frame all of it around chat, so racing video models meant scrolling past a question console and an app builder to find them. A switch at the top now asks: All, Chat and writing, Code, Makes images, Reads images, Makes video, Listens to audio or Makes audio. The Models screen, the Lab, Comparison and Scorecards follow it, and each use crowns its own winner from its own measurement. It opens on what you chose when you first ran RigMatch, and remembers what you pick.
• RigMatch can race video models on your own machine. The catalog now holds seventeen text-to-video models, each downloaded only if you pick it. The ones you pick render the same prompt with the same seed, one after another, fastest first, with ComfyUI unloaded between them once you agree, so none inherits a warm start. Their frames are judged only after the last one finishes, because the judge stays in graphics memory for ten minutes after it answers and would otherwise be timed as part of the next model. The race ends as a leaderboard with the clips side by side, and a model that fails says why while the rest carry on. On an RTX 4070, three Wan models took 2 minutes 41, 2 minutes 41 and just under 6 minutes.
• Before you download a video model, RigMatch says whether it will run here and how long a clip should take. The lineup runs from seventeen seconds to three hours a clip and up to 56 GB of download, and neither answer is on a model's download page, because both depend on your machine. Each model says it fits in graphics memory, fits by borrowing system memory, is too big, or needs a Hugging Face token, with a time range for this computer, and the size shown is everything it needs rather than its main file alone. Your first LTX-Video 2B render calibrates the times to your card, and a time measured here replaces the estimate outright. The check is not right for every model yet: Mochi 1 was shown as fitting a 12 GB card and ran out of graphics memory at its last step, 24½ minutes in.
• Big downloads survive a dropped connection. A 40 GB file used to be deleted and started again from nothing; it now resumes where it stopped, and it has to match the catalog's size and the checksum Hugging Face publishes before ComfyUI is allowed to see it. Stop still throws the partial file away, because nobody asked to keep 40 GB of fragment. Every site a download is redirected through is checked against the ones RigMatch trusts, not only the first. Some models need a Hugging Face account: Settings takes a token, and it is only ever sent to Hugging Face. Where a model's license limits who may use it, the download dialog quotes that condition before you agree.
• RigMatch can make audio. Makes audio runs three models on your own ComfyUI, adding nothing to it: ACE-Step 1.5 Turbo (10.03 GB), ACE-Step v1 3.5B (7.70 GB) and Stable Audio Open 1.0 (4.85 GB, plus the 0.89 GB encoder that reads its prompt). Each gets the same prompt, seed and 30-second length, from a model's own row or several at once on Comparison, and every clip plays where it lands. On an RTX 4070, ACE-Step 1.5 Turbo made 30 seconds of dance music in 8 to 11 seconds, and Stable Audio Open made 30 seconds of rain in 7 to 8. Stable Audio Open's community license condition is shown before it downloads.
• A model that can hear was meant to judge each clip, and the Gemma 4 listeners in Ollama cannot tell rain from music. Asked about real clips, both answered No to every question, and one took the rain for a barking dog. Every benchmark prompt mixes questions whose answer is yes with ones whose answer is no, so a listener that gives them all the same answer has not told one clip from another: RigMatch now leaves that clip unjudged instead of scoring it 33%. Your ear is the judge instead. Under each clip, Sounds right or Doesn’t becomes its accuracy, marked as yours wherever it is ranked, and pressing it again takes it back. It works for a prompt you wrote yourself, which nothing else can check.
• Comparison follows what you are testing. For chat, code and reading pictures it is Speed Dating, ranked at that use's fader. Pictures, video and audio are made rather than answered, so there it runs the models itself: tick the ones that can run here, pick a prompt, and they render one after another with the same seed, every one finished before any is checked. What each made lands side by side, and only results given the same prompt are ranked together, because a model that drew the lighthouse it was asked for is not beaten by one that drew a cat more quickly. Scorecards follows the same choice, with every saved result dated and ranked at the fader.
• A picture or video model's row now says Test where it said Open Lab, and an audio model's row has one too. Open Lab sent you to another screen to pick the model you were already looking at from a second list. The test opens under the row instead: a prompt, the fader, Run and the latest result. It runs exactly what the Lab or Comparison would run and saves where they save, so every screen sees it. A row's Test also follows what you are testing: on Listens to audio it plays the listening passage, and on Reads images it shows the model a test picture, where both used to ask the chat questions.
• You can see what is rendering from any screen. A test keeps running when you close the panel that started it, and nothing said so: a video started from its row could render for twenty-five minutes and fail with no sign anywhere. The status bar now shows what ComfyUI is making, for which model and for how long, with a Stop button. Activity reads Live in the side menu, the model's row reads Testing, and Activity has a Renders card. When a render ends, its verdict stays in view until the next one, and one you stopped says it was stopped rather than that it failed.
• ComfyUI starts itself when a test needs it. Choosing Makes images, Makes video or Makes audio starts it, and so does opening a model's Test while it is down. It uses the launcher beside the ComfyUI folder you gave RigMatch, the file you would double-click, and runs nothing it did not find there. It starts once a session, so if you close ComfyUI it stays closed, and a start that never answered waits for you to press Start. Settings can turn this off. ComfyUI keeps running after RigMatch closes, as it would if you had started it by hand.
• Simple Mode races video makers too. Choosing "a video maker" used to end in a paragraph sending you to Advanced Mode. The race now runs in place, in a simpler form: only the models that can run on this computer, a button that picks the fastest three, and the same leaderboard with the clips side by side. The count on that choice also stopped including models too big for your machine, for pictures as well as video.
• Judging is more dependable. Tests picked their judge from the whole catalog, so a vision model you had never downloaded could be chosen, every question to it failed, and results came back unjudged while the fader said a judge was there; they use one you have installed now. A thinking model asked to judge a frame spent its whole reply thinking and answered nothing; it is now asked not to think first. And a line nothing could check, such as a picture no judge could answer about, or a video's motion, which nothing measures, read Miss beside a board that called it unjudged. It reads Not checked now.
• Speed is measured with the computer to itself again. Chat and writing answers are marked by a judge, and the judge was the largest text model you had installed, whether or not it fit, kept loaded for ten minutes in the middle of the timing. On an 8 GB Jetson that was a 12 GB gemma4 running mostly on the CPU, the model under test was pushed off the graphics card, and qwen2.5:1.5b scored 9 for speed. The judge is now one that fits this computer, it leaves as soon as it has answered, and the model is loaded again before its next timed answer. The same test then scored 26 and finished in 3 minutes instead of 9½. If you tested on a computer with little graphics memory, test again, because those speed scores came out too low.
• On a Mac, the RigMatch Chat in the disk image is RigMatch Chat. In every Mac release until now it was a second copy of RigMatch under Chat's name, which is why the download was 250 MB when the app itself zips to 125. If you dragged it to Applications, drag the new one across and choose Replace. The Chat button inside RigMatch opens the real one either way.
• Reads images now lists every model that can see. It went by the catalog's descriptions rather than asking Ollama, so gemma4, described for quick chat, was missing while Ollama said it could read pictures. Earlier versions offered a Wan 2.1 1.3B download that could ...

v0.8.0-beta — It Says What It Measured

What's new in 0.8.0 — It Says What It Measured

• There is a new set of questions about difficult subjects — documented history and openly disputed questions, to see what a model will discuss and whether it gets it right. It began as a "Censorship" test and is not called that any more, because that is not what it turned out to measure: a model can be perfectly willing to answer and still be wrong, and willingness alone was never the interesting part.
• Grading those answers is not something this computer can do alone, and RigMatch no longer pretends otherwise. The local scorer can tell a refusal from an engaged answer; it cannot tell a complete answer from a confident, fluent, incomplete one. Those questions now say "needs a judge" rather than producing a number that looks like a verdict, and where a judge has graded them the run records that it did.
• The judge was also taught that leaving something out is a way of being wrong. An answer can be accurate in every sentence and still omit the central documented fact — describing Tiananmen Square in June 1989 as a demonstration that was brought to an end, without mentioning that people died. The rubric now grades completeness alongside accuracy and willingness, and deliberately does not require any particular figure or wording, because the point is whether the answer is whole rather than whether it matches a script.
• Before a run starts, RigMatch says what that run would send off this computer. Its premise is that nothing leaves, and for a local model graded by the local judge that is exactly true. Two settings break it: a cloud model answers remotely, and the OpenRouter judge is sent both the question and the answer to grade it. Those are named outright before you start. Whether that matters is your context to weigh — where you are, what you are asking about — and the app states what happens rather than assuming an answer for you.
• A finished comparison now tells you it finished, and shows you the report where you are standing. The results used to land in three places and announce themselves in none of them: the ranking in one screen, the answers in another, the winner in a badge. There is a report you can open the moment a run ends, holding the ranking, what each model turned out to be best at, and how they each answered the same question side by side. Recent reports are kept, and Activity lists them.
• The Comparison screen now leads with the answer. Its parts are a menu rather than a stack of cards read through a slot, and the ranking is the first thing in it — the screen was opening on the details and keeping the result below them. The ranking also says what it is a ranking of: crowning a best match out of three models when five are in your lineup is a different claim from crowning one out of five, and the screen used to make both in the same words.
• Underneath the ranking is what each model is actually best at. Every run has measured this since task groups arrived and no screen had ever drawn it, so "which of these three is best at coding" was being answered on every comparison and shown nowhere. A column only names a winner where two or more models were measured well enough to have one, and a tie is reported as a tie.
• Two models can also be compared directly, with the same restraint: an untested model is unmeasured rather than losing, a gap of a point or two is reported as a gap a re-run could close, and size and compression are described without declaring a winner, because neither is better in the abstract.
• The Models screen was rebuilt around finding things. The filter tray is gone, replaced by rails that are always visible and say what is active; search suggests what you might mean; and families of one model in twelve sizes collapse to a single row you can open. Every model carries a face, and where the organisation behind it is known, the country it comes from — as its own column and its own filter. Where the country is not known, nothing is guessed.
• A collapsed family now puts its best version forward instead of its first. Two of three families on a real machine were advertising their weakest member: Qwen2.5, which had scored 92 on its 7B, introduced itself as a 0.4 GB model. Sorting the table also genuinely reverses now — the tie-break used to run the same way whichever direction the column pointed, and on a list where most rows tie, that was most of the ordering.
• The Popularity column had quietly stopped being one. Ollama changed the markup of its library listing, the download counts all came back empty, and rather than failing the column renamed itself and showed speed instead — a whole column changing meaning with nothing anywhere reporting a problem. It reads download counts again, and the release checks now confirm against the live site that every model in the catalogue still carries one.
• Simple Mode was measured rather than admired. Its first button sat below the fold on an ordinary laptop; the model cards were so tall that one row of contestants filled the entire step and two-thirds of the choice was off screen; and the finished Compare step looked complete but did nothing when clicked. The cards lie across now instead of down, the first thing to press is the first thing you see, and the Compare step opens the report — which was previously unreachable from Simple Mode at all. Each card also shows how many people have downloaded that model, which is the one signal a beginner can read without knowing what a parameter is.
• Simple Mode stopped captioning a question about the Tiananmen Square massacre as "Everyday questions". It was guessing what each question tested by pattern-matching its title, and every question in the difficult subjects set is titled by its subject rather than its kind, so all of them fell through to the blandest label available. The question knew what it was the whole time; the app simply was not passing it along.
• The Top Match badge says what the model is top for, and is no longer cut off saying it. Setup also stopped calling a computer ready for something it cannot do, and Check Local now shows that it ran and what it found rather than finishing in silence.

Downloads

• Windows: RigMatch-*-win-x64.exe
• macOS Apple Silicon: RigMatch-*-mac-arm64.dmg
• macOS Intel: RigMatch-*-mac-x64.dmg
• Linux x64: RigMatch-*-linux-x86_64.AppImage or RigMatch-*-linux-amd64.deb
• Linux ARM64 / Jetson: RigMatch-*-linux-arm64.AppImage or RigMatch-*-linux-arm64.deb

Already running an older install? It updates itself to this version automatically — nothing to re-download.

Windows first launch

RigMatch for Windows is an unsigned beta, so Microsoft Defender SmartScreen will almost certainly stop it the first time. That warning means the file has no code-signing certificate yet and has not been downloaded enough times to build a reputation — not that anything is wrong with it.

1. Download RigMatch-*-win-x64.exe.
2. If SmartScreen says "Windows protected your PC", choose More info, then Run anyway.
3. If your browser blocks the download instead, choose Keep and, where asked, Keep anyway.

Prefer not to click through a warning? Download the .zip instead, unpack it anywhere, and run RigMatch.exe from the folder — a portable copy that installs nothing.

Either way, verify what you downloaded against SHA256SUMS.txt on this release:

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

A signing certificate is on the roadmap; until it is bought and the reputation builds, this warning is expected on every release.

macOS first launch

RigMatch for macOS is currently an unsigned beta distributed outside the App Store. On first launch, macOS may say the developer cannot be verified or that the app was downloaded from the internet.

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

After that first approval, RigMatch opens normally by double-clicking. If macOS says the app is damaged after copying it to Applications, run this Terminal command once:

xattr -cr /Applications/RigMatch.app

Apple reference: https://support.apple.com/guide/mac-help/open-a-mac-app-from-an-unknown-developer-mh40616/mac

Full Changelog: v0.7.1-beta...v0.8.0-beta

v0.7.1-beta — It Fits The Machine You Have

What's new in 0.7.1 — It Fits The Machine You Have

• A Jetson's graphics is not a card on the PCI bus, so RigMatch's hardware scan came back empty. No name, no memory. It then did what it does for a machine with no graphics at all: reported 0 GB, recommended only the smallest models, and announced that no NVIDIA GPU was present, on NVIDIA hardware. It now reads the board's own name from the device tree and recognises the shared memory pool for what it is. On an Orin Nano that took the reported memory from 0 GB to 7.4 GB, and the models it considered usable from 5 to 140.
• The window would not go smaller than 1280 by 820. That is fine until you meet a 1366x768 laptop, the commodity panel for a decade, or a 720p monitor: the window opened taller than the screen, and the bottom of the app could not be reached at all. What lives down there is the ticker, including the button that opens RigMatch Chat. Dragging it up did not help either, because no legal size existed for the window manager to move it to. The floor is now 1024 by 640, which those machines fit with room for a taskbar.
• Two layout rules met on a window neither expected. One hides the number badge when the window is short, the other hides the icon when it is narrow, and on a window that was both, every destination in the sidebar collapsed to “M..”, “W..”, “C..”: present, in the right order, and unreadable. It took a window smaller than RigMatch used to allow, which is why nobody had seen it.
• Choosing “create an image from a prompt” looked no different from choosing everyday chat, and nothing said it needed a second program. Three steps later the download refused and told you to open Settings, a panel Simple Mode does not have, so the only way forward was to find Advanced Mode. Goals that run through ComfyUI now say so on the tile before you pick one, and the refusal offers to go and find ComfyUI rather than naming a place you cannot reach. If it simply is not running yet, RigMatch says that and keeps the button, so starting it and pressing again is the whole fix.
• RigMatch Chat has its icon back on macOS. Dragging only RigMatch out of the disk image, which is the ordinary thing to do, left the companion as a bare program file rather than an app, and macOS gives those a blank icon and no name in the Dock. RigMatch now carries the whole companion app inside itself.
• The Mac and Linux downloads are about 15 MB smaller. Every package was carrying the Windows companion alongside its own, an executable those systems have no way to run.

Downloads

• Windows: RigMatch-*-win-x64.exe
• macOS Apple Silicon: RigMatch-*-mac-arm64.dmg
• macOS Intel: RigMatch-*-mac-x64.dmg
• Linux x64: RigMatch-*-linux-x86_64.AppImage or RigMatch-*-linux-amd64.deb
• Linux ARM64 / Jetson: RigMatch-*-linux-arm64.AppImage or RigMatch-*-linux-arm64.deb

Already running an older install? It updates itself to this version automatically — nothing to re-download.

Windows first launch

RigMatch for Windows is an unsigned beta, so Microsoft Defender SmartScreen will almost certainly stop it the first time. That warning means the file has no code-signing certificate yet and has not been downloaded enough times to build a reputation — not that anything is wrong with it.

1. Download RigMatch-*-win-x64.exe.
2. If SmartScreen says "Windows protected your PC", choose More info, then Run anyway.
3. If your browser blocks the download instead, choose Keep and, where asked, Keep anyway.

Prefer not to click through a warning? Download the .zip instead, unpack it anywhere, and run RigMatch.exe from the folder — a portable copy that installs nothing.

Either way, verify what you downloaded against SHA256SUMS.txt on this release:

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

A signing certificate is on the roadmap; until it is bought and the reputation builds, this warning is expected on every release.

macOS first launch

RigMatch for macOS is currently an unsigned beta distributed outside the App Store. On first launch, macOS may say the developer cannot be verified or that the app was downloaded from the internet.

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

After that first approval, RigMatch opens normally by double-clicking. If macOS says the app is damaged after copying it to Applications, run this Terminal command once:

xattr -cr /Applications/RigMatch.app

Apple reference: https://support.apple.com/guide/mac-help/open-a-mac-app-from-an-unknown-developer-mh40616/mac

Full Changelog: v0.7.0-beta...v0.7.1-beta

v0.7.0-beta — It Tells You What It Cannot Do

What's new in 0.7.0 — It Tells You What It Cannot Do

• The chat no longer pretends. Ask a local model to draw you something and it will not refuse — it will describe a picture warmly, or announce that it has made one, and until now RigMatch passed that on without comment. Ollama serves text; none of these models can produce an image, a video or a sound. RigMatch now says so in its own voice before the model answers, names the model that cannot do it, and points at the place in the app where the thing can actually be done. The same goes for asking a text model to transcribe a recording, which it will answer with a plausible transcript of nothing at all — the most convincing of these untruths, because there is no missing picture to give it away. Models that genuinely hear are not warned about, because some of them really can.
• And where this computer can do it, RigMatch offers to. If ComfyUI is running with a model that can draw, asking the chat for a picture turns the note into a button: press it and the image arrives in the conversation. Your own words are the prompt, not a rewritten version of them, because a picture of a request you did not make is worse than no picture. It reports elapsed seconds rather than a progress bar — ComfyUI says nothing between starting and finishing, so a bar would be invented — and Stop genuinely stops it. Video is deliberately not offered here: a few seconds of footage is megabytes, and a conversation is not the place for it. The Video test in Advanced Mode remains where that belongs.
• RigMatch Chat stops making you deduce what your models can do. The sidebar asks "What do you want to do?" — write, read a picture, hear audio, make a picture — and the buddy list answers by narrowing to the models that genuinely can, with sees and hears marked on the buddies themselves. Choose "make a picture" and every chat model drops out, because the honest answer to which model makes an image is none of them: what appears instead is the image maker itself, named — the actual ComfyUI checkpoint that would do the work — with Ready or Not ready stated outright. Open it and you get a workspace rather than a conversation: describe the picture, press Make image, watch elapsed seconds, and the file lands in your Pictures folder with its path shown. It is deliberately not a chat, because there is no model on the other side and nothing should pretend to talk back.
• That Ready is now kept true minute by minute. RigMatch used to look for ComfyUI once, at startup — so starting RigMatch first, which is the normal order, meant Not ready for the whole session even with ComfyUI running and a checkpoint loaded, and closing ComfyUI mid-session left the offer standing when it could no longer be honoured. It now looks every fifteen seconds: stop ComfyUI and the offer withdraws within seconds, start it and the offer returns with the checkpoint named.
• Two RigMatch windows can no longer quietly disagree about who is answering. RigMatch Chat talks to whichever RigMatch owns the local bridge, and a second window used to lose that race in silence — the companion would list one window's models and save the other window's pictures, with nothing anywhere saying so. Launching a second RigMatch now brings the first one forward instead, and a window that somehow does not own the bridge refuses to open the companion, saying another RigMatch is already running rather than launching something wired to the wrong app.
• A security pass on that bridge, because more people are about to run this. The local port RigMatch answers on accepted requests that carried no browser origin at all, which meant anything on your machine able to open a socket could read your model list and start graphics card work. It now refuses to start work for anyone but RigMatch Chat, and refuses any request whose Host header is not local — the DNS rebinding trick, where a web page points its own name at your machine. Every one of those rules is exercised against the real running app by the release checks, not just written down. Every source file also now carries a copyright line.
• The listening test shows you the microphone. Nothing on that panel gave any sign it was picking anything up — you read a script into it, pressed Run, and found out afterwards. There is a level bar while recording now, and it moves with your voice. More importantly, a silent recording is no longer scored: sending silence to a model and reporting that it scored zero for hearing nothing was true twice over and a lie once, because the fault was a microphone and the number blamed the model. RigMatch also tells the two apart — the meter watches the microphone before RigMatch touches the audio, so if it was picking you up and the recording still arrived empty, that is this app losing it, and it says so rather than sending you to check hardware that was working.
• Image models are asked for what they were built for. Every checkpoint was being run at twenty steps with guidance at seven, which is right for ordinary Stable Diffusion and wrong for a distilled model like SDXL-Turbo — built to finish in a handful of steps with guidance off, and pushed to seven it returns something oversaturated and posterised. The benchmark then marked the model down for how well that matched the prompt, which was a request RigMatch had made badly rather than a failure of the model. Steps and guidance now come from the checkpoint, and anything unrecognised keeps the settings it always had. Image scores saved before this were measured the old way and are not comparable with new ones.
• ComfyUI can be started from RigMatch. It is a separate program you run yourself, which until now meant leaving the app to go and find a batch file — even though RigMatch already knew where the install was, having checked that folder to put downloads in. The button sits in the status strip at the top, where you notice ComfyUI is not running, and it only appears when there is genuinely a launcher there to run. It says ComfyUI is starting rather than ready, because loading takes a moment and only the status check can honestly say when it is done.
• A score now records the computer that earned it. RigMatch can benchmark a model on an Ollama running on another machine on your network, and until now the result came back stamped with this computer's graphics card, memory and driver — the run had happened somewhere else entirely. It cannot ask a remote machine what card it has, so it now records that machine's name and no hardware at all, which is the honest answer. Those scores are marked "Measured on another computer — retest here" and do not crown a winner, for the same reason a score from a graphics card you have since replaced does not. Testing on the computer in front of you, which is what nearly everyone does, is unchanged.
• RigMatch now checks the graphics card before the Advanced Lab runs, not just before a benchmark. A ten second recording sent to a listening model while a game held the card took over four minutes and then reported a connection timeout — the service was reachable the whole time and simply starved. The lab says the card is busy before you start, the chat says it while a picture is being made, and the reading is taken before RigMatch adds any load of its own, so it never blames you for work it is doing itself.
• Timeouts stop blaming the connection. "Timed out reaching local AI service" reads as "Ollama is down" and sends you to check a service that was answering all along. It now distinguishes the two: the request was accepted and did not finish in time, the service is reachable, and a busy graphics card or an oversized model is what does this.
• Clear Data now clears your data. It removed six saved settings while the app was writing twenty-four, and then reported "RigMatch app data cleared". Left behind were your model notes, your run history, your goals, your ComfyUI folder, your grading setup — and any OpenRouter API key you had saved, which is the one that matters: clearing your data before passing a laptop on should take a stored credential with it. Everything goes now, and both dialogs say what is kept.
• What is kept is your Simple or Advanced choice, and the fact that you have already read the getting started guide. Clearing your scores used to demote an Advanced user to Simple Mode and replay the tour, which is not what anyone means by clearing data. Your goals are genuinely cleared, so you are asked for them again — that is the honest consequence rather than an oversight.
• It was also possible for Clear Data to do nothing at all and say so unhelpfully. The run log is cleared first and is rate limited, so clearing your logs and then clearing your data a moment later threw an error and abandoned the whole thing, leaving everything in place behind the words "Could not clear all data". The log clear is now reported alongside the wipe rather than instead of it.
• The listening test lets you choose which microphone. It took whatever Windows called the default, which on a machine with a headset, a webcam and a USB interface is a coin toss you cannot see. The record card also now says to read the script word for word — improvising marks the model down for hearing you correctly, and the panel had no way to explain that.
• An attachment can no longer outlive the model that could read it. Recording something for a model that hears and then switching to one that does not left the recording attached, and sending it produced "Failed to load image or audio file" from Ollama — which reads as a broken recording rather than the wrong model. It is now removed with a sentence saying why. The chat model list also stopped offering image generators as things to hold a conversation with.
• Six labels on the Models screen were too small to read, and are not any more. That single fix also cleared the layout check across all sixteen window sizes and modes, where those labels had been the only complaint.

Downl...

v0.6.0-beta — What Do You Want To Do?

What's new in 0.6.0 — What Do You Want To Do?

• RigMatch now asks what you actually want before it asks anything else. The first screen used to be "Simple or Advanced?" — a question about our software. It now opens with twelve things a person might want a model for, grouped as Chat, Work, Image, Audio and Video: talk to a model, help me write, help me code, transcribe a recording, describe a picture, make an image, make a video, power my automations. Pick one and the app points itself at that: the model list opens already filtered to models that can do it, and Simple Mode opens on the same choice. You can pick more than one, but it gently talks you out of it, and says why — the best model for coding is rarely the best for chatting, and every goal you add spreads your testing time thinner.
• Each goal tells you what to expect from your hardware before you spend an hour finding out. Video generation on a 12 GB card reads "A good match for your rig" with the measurement behind it — about twelve seconds of work for four seconds of 768p video, measured, not guessed. On a smaller card the same goal says it might be out of your league, and explains that the video models alone want about 11 GB. Where the number comes from a real run it says so; where it is a rule of thumb it says that instead, and the whole app keeps insisting that the test is the real answer.
• There is more than one winner now. Asking "which model won?" makes no sense if you picked coding and image-making — a coder and a painter cannot lose to each other. Scorecards now crowns one model per goal you chose, with your first pick headlining as Your Match and the rest alongside it. A crown only ever comes from questions actually asked on your machine; where nothing has measured a goal yet, the card says exactly what is missing rather than quietly borrowing a number from somewhere else.
• RigMatch can make pictures and video, not just talk about them. Ollama cannot run image or video models, so this works through ComfyUI — a separate free program the app does not install or bundle. Point RigMatch at your ComfyUI folder and it can download a curated set of image and video models straight into the right places, then time them properly: an image model is scored on how fast it renders and whether the picture matches the prompt, judged by a model that can actually see. Video is scored on speed per second of footage and whether the frames match what you asked for, and the app says outright that it cannot yet measure whether the motion looks good.
• The folder check is deliberately paranoid, because two ComfyUI installations on one machine is normal and multi-gigabyte downloads landing in the wrong one is not. RigMatch compares what is on disk against what the running program reports and refuses the folder outright if they disagree.
• Answer quality is measured more honestly, and in one place it now refuses to answer at all. RigMatch grades every answer by matching a shape — valid JSON with the right keys, a refusal where one belongs, a list of the length you asked for. Everyday chat and writing have no such shape, and what the app had been doing instead was scoring those answers by their LENGTH: a fixed number for any reply, rising with the character count. That is not a measurement of quality, and it was being used to crown "Best for talking" — so the wordiest model won. It no longer crowns anything. Instead, RigMatch quietly hands those particular answers to a second model you already have installed, which reads them and marks them properly — so chat and writing get a real score out of the box. Only the questions that need it are marked this way, so a run of JSON or automation questions costs nothing extra and a chat run pays only for its chat questions. It is always a model on your own machine, never a paid cloud one. Switching the judge on in Settings still means judge everything, and if nothing you have installed can grade prose the app says so rather than inventing a number. Everything the rules can genuinely grade — JSON, refusals, formatting, and the one coding question with a known right answer — is unchanged.
• Because of that, saved scores from before this version are marked for a retest, and older scorecards will show "Retest recommended" until you run them again. The way answers are grouped changed underneath them: the JSON questions used to be pooled with formatting as "does it do as it is told", and they are now counted on their own, because that is what they measure and it is what crowns the automations goal. A score built the old way and one built the new way are not the same measurement, so RigMatch will not rank them against each other.
• Writing is a real goal now rather than a promise. RigMatch shipped a Writing test whose questions were filed as ordinary chat, so they were scored as chat and the goal they belonged to could never be crowned. They are filed correctly, and writing is graded on them — marked by a second model, as above, so the goal crowns a winner without you configuring anything.
• Each goal now knows which test measures it. Pick "help me code" and the Coding focus is flagged as yours in the test dialog, one click away, so a shorter run still asks enough coding questions to name a winner. There is a new Tools & Automations focus as well, which was the one goal with nowhere to send people; it crowns a winner in a single ten-question run.
• Saved scores now remember the computer that produced them. A score is only ever true of a rig, and both halves of that can move underneath it — you change a graphics card, or a model tag quietly updates to different weights under the same name, which Ollama does routinely. Every new score records the graphics card, its memory, the driver, the app version and the exact fingerprint of the weights tested. If any of that no longer matches, the score is badged for a retest and it stops crowning winners, because a number measured on a different setup should not be picking your match.
• Every Run button now tells you roughly how long it will take, learned from your own past runs on this machine rather than a fixed table. Runs from a different computer are ignored, because a laptop's pace says nothing about the desktop its history moved to.
• Settings has a Closet: every model you have installed, biggest first, with how much disk it takes, what it scored, and whether it has ever been tested at all. Your current top match wears a crown; everything else offers to be evicted, one at a time, with the score in front of you — the calm version of the clean-up prompt that used to appear on the way out the door.
• You can save a match card. Every result until now lived and died in browser storage on one computer, with nothing to show anyone. The winner can now be saved as a picture — score, grade, and the graphics card that measured it, because a score without the rig invites exactly the comparison this app spends its time refusing to fake. It is drawn on your machine and saved as a file; nothing is uploaded anywhere.
• Models that cannot do the same things are no longer put in the same comparison. An image generator and a chat model scored against the same questions produces a meaningless ranking with a guaranteed loser, so a lineup now has to share the ability it is being graded on.
• The listening test takes a recording or your microphone, and the model list can be filtered by what models can do — hears audio, watches video, makes images, makes video — including models you have not downloaded yet, so the filter is a way to go shopping rather than just a way to sort what you already own.
• Under the hood: a test now checks every color pairing in all five themes against the readability standard, so a new theme cannot ship unreadable text; and two hundred colors that had been written in by hand now come from the theme.
• Simple Mode explains itself now. The host does it: point at any word you do not know — graphics card, Ollama, VRAM, Match Score, download size — and he stops narrating and explains it instead, in one sentence, without leaving the screen or opening anything. The first screen also now says what an AI model actually is and what you get out of this, before it asks to look at your computer; it used to open with a hardware scan for a thing it had never named.
• The Winner screen shows the whole result, not just the winner. It used to announce one model “out of the 5 you tested” and then show nothing whatever about the other four — the comparison you waited for was reduced to a single number. There is now a board with every model that ran, in order, with its score. A close second may suit you better, and now you can see there was one.
• The comparison screen tells you how much longer it will take. It used to show only a question count, which is not much comfort a quarter of an hour in. It now times the run in progress and says “about 5 minutes left”, and shows each answer's score as it lands, so you can see how a model is doing rather than only that it is busy.
• When RigMatch asks you to review a model's terms before downloading, it now links that model's terms. It previously showed the same fixed list every time — so downloading DeepSeek presented Gemma's licence and prohibited-use policy, three documents that did not apply to it and none that did, on the one screen whose whole job is making sure you know what you are agreeing to.
• The app no longer claims a previous show's winner is leading the current one. Swapping a single contestant was enough to make the comparison screen announce a model that was not in the lineup at all as the one holding the top score. Results from an earlier lineup are now labelled as exactly that.
• Fixes for smaller and shorter windows. On a laptop screen the menu quietly cut Activity and Settings off the end of the list, so Settings looked like it did not exist; and in a window narrower than about 1100 pixels the Advanced screen sliced...

v0.5.0-beta — It Can Hear

What's new in 0.5.0 — It Can Hear

• RigMatch can test whether a model can listen. Some local models take audio as well as text, and nothing in the app knew it — one already installed on a typical machine can transcribe speech and had never been asked to. The new Listening test plays a short spoken passage and compares what the model wrote down against the words that were actually said, word for word.
• This is the first score in RigMatch measured against a right answer. Every other quality number here is a judgement — a set of pattern checks, or a second model marking the first one's homework — and the app has always been honest that those are a rough guide. A transcript can simply be right or wrong, so the listening score is the real thing: a model that hears nine tenths of the passage scores 90, and that means exactly what it says.
• You can also send a recording in chat, to any model that can take one. Attach it the same way you would a picture, and it plays back in the conversation so you can tell at a glance which file you sent.
• The app now asks each model what it can do rather than guessing from its name. Ollama reports this directly — whether a model can hold a conversation, read pictures, hear audio, or generate images — and RigMatch reads it. That fixes a handful of quiet mistakes: models were being labelled by keywords in their names, so anything published in Ollama's community area was tagged as an image generator whether it was one or not.
• It also stops a model being set up to fail. Ollama can download image-generation models but cannot actually run them yet, so putting one in a comparison guaranteed it a bottom score for something that was never its fault. Those are now kept out of comparisons and chat until they can genuinely run.
• The "Makes video" filter is hidden, because nothing can satisfy it. There is no video generation on Ollama at all — everything that looks like it is a model that watches video rather than making it. The filter reappears by itself if that ever changes.

Downloads

• Windows: RigMatch-*-win-x64.exe
• macOS Apple Silicon: RigMatch-*-mac-arm64.dmg
• macOS Intel: RigMatch-*-mac-x64.dmg
• Linux x64: RigMatch-*-linux-x86_64.AppImage or RigMatch-*-linux-amd64.deb
• Linux ARM64 / Jetson: RigMatch-*-linux-arm64.AppImage or RigMatch-*-linux-arm64.deb

Already running an older install? It updates itself to this version automatically — nothing to re-download.

macOS first launch

RigMatch for macOS is currently an unsigned beta distributed outside the App Store. On first launch, macOS may say the developer cannot be verified or that the app was downloaded from the internet.

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

After that first approval, RigMatch opens normally by double-clicking. If macOS says the app is damaged after copying it to Applications, run this Terminal command once:

xattr -cr /Applications/RigMatch.app

Apple reference: https://support.apple.com/guide/mac-help/open-a-mac-app-from-an-unknown-developer-mh40616/mac

What's Changed

• 0.5.0 — It Can Hear by @DaveEuson in #16

Full Changelog: v0.4.4-beta...v0.5.0-beta

v0.4.4-beta — Readable, and Reachable

What's new in 0.4.4 — Readable, and Reachable

• Starting a show from Simple Mode can no longer fail with a programmer's error message. If none of the models in your lineup could actually run — which can happen even after they have all downloaded, because a model also has to suit your operating system — the show started anyway with nobody in it and fell over on the way to the results, showing you the words "Reduce of empty array with no initial value". Simple Mode now checks the lineup the same way Advanced Mode always has, and if there are fewer than two models that can run it says so plainly and lets you go back and pick more.
• The red and blue status text throughout the app is now bright enough to read. Both were set at a strength meant for a light background and never adjusted for a dark one, so warnings, "out of your league" tags, size labels and error notes were low-contrast grey-on-dark in all five themes — worst on the raised panels, which is exactly where most of them appear. They are the same colours, just lifted until they clear the standard for readable text. Five more spots that had their colour written in by hand rather than taken from the theme were fixed too, including the remove button on a model card and the placeholder in the notes box, both of which were close to invisible.
• Dialogs now behave like dialogs if you use a keyboard or a screen reader. Twenty windows in the app announced themselves as ones that block the page behind them, and none of them did: focus started outside the window, Tab walked straight out of it into the page underneath, and closing it left you wherever you had wandered to. Only four could be dismissed with Escape. Now the keyboard goes into a dialog when it opens, stays inside it, returns where it came from when it closes, and Escape works everywhere it should. The two that delete things — removing a model and clearing your data — were the worst of these: no Escape, no click-outside, so the only way to back out was to hunt for the Cancel button by tabbing through the page behind it.
• Four things that called themselves dialogs were not, and pretended less rather than more: the update notice in the corner is a status message, the welcome tour is a coach mark you are meant to click past, and the test editor panel leaves the app usable behind it. Trapping the keyboard inside any of those would have been the wrong fix.
• RigMatch Chat now remembers your conversation properly. It had been asking every model to hold about 4,000 words at a time — Ollama's fallback, not the model's own limit — even for models built to hold thirty times that. Past that point the oldest messages quietly stopped existing as far as the model was concerned, while they stayed on screen in front of you, so a perfectly good model would answer "this conversation just started" about something you had said two messages earlier. Each model is now asked for as much as it can genuinely hold on your hardware — RigMatch Chat reads how much video memory your card actually has and sizes the window to it, taking the model's own weights into account, since both come out of the same pool. On a 12 GB card that works out at two to eight times more memory across a typical set of installed models.
• There is also a memory gauge next to the model name, so you can see how much room is left in the conversation before anything starts dropping out. It turns amber as it fills and says so outright when it is full. If you would rather trade video memory for a longer memory, or the other way round, Settings now has the choice — with what each size costs shown next to it, and a note on any size too big for your card to hold, which would push the model onto the processor and make it crawl.
• Chat history is no longer capped, and no longer stutters. Conversations were kept in browser storage, which stops at about 5 MB — and RigMatch Chat was rewriting your entire history, every conversation with every model, once for each word as it appeared on screen. On a well-used history that worked out to roughly 3 GB written and two and a half seconds of frozen window for a single reply, and once the 5 MB was reached the app crashed to an error screen on startup and did it again every time you opened it, with no way back. History now lives in a normal file with no size limit, and is saved a handful of times per reply instead of hundreds. Anything already saved moves across by itself the first time you open this version. If saving ever does fail, the app says so in a bar at the top rather than falling over.
• You can stop a reply now. There was no way to interrupt one — the box you type in simply locked until the model finished, however long that took and however wrong the answer was going in. Send becomes Stop while a reply is coming, and pressing it genuinely stops the work rather than just hiding the rest: the graphics card is released and whatever had already arrived stays on screen, because a half-answer is often all you wanted. You can also type your next message while a reply is still arriving, which you could not before.
• RigMatch Chat can hold more than one conversation with a model. Each model had exactly one thread, forever — no way to start a second subject without losing the first, and nothing to tell them apart if you could. Every model in the list now folds open to show its conversations, named automatically after the first thing you asked in them, newest at the top. You can start a new one, rename any of them by double-clicking, and delete the ones you are done with. Everything you already have is carried across and given a name on first launch.
• Switching personality no longer looks like it deleted your chat. The personality was quietly part of how a conversation was filed, so changing it swapped you to a different, usually empty thread — with nothing on screen to say why, and no obvious way back. It now belongs to the conversation: change it and the chat in front of you stays put, and a thread you started as Creative Copilot is still Creative Copilot when you return to it next week.
• When a chat fills up the model's memory, you can now do something about it instead of watching it quietly forget. RigMatch Chat offers to summarise the older part of the conversation and send those notes in place of it — measured on a real chat, that cut what the model had to read by about three quarters while keeping every fact in it. You can compact the chat you are in, or start a fresh one that carries the summary across and leaves the original untouched. Either way nothing is deleted: the whole conversation stays on screen, with a line showing exactly where the model's memory begins, and you can read the notes it is working from at any time.
• The summary is written by whichever of your models answers most accurately, which is not always the one you are chatting with — and not always the one with the highest overall score, since that score rewards speed and the fastest model is usually the smallest. It shows you the notes before using them, and you can edit them first, because a summary that quietly gets something wrong would colour everything the model says afterwards.
• Category picks like "Best for coding" are now measured on your machine rather than looked up. RigMatch has always asked five different kinds of question during a benchmark and scored every answer separately, but it averaged them into one number and threw the breakdown away — so "Best for coding" actually meant "whichever model a built-in list describes as a coding model happens to score highest overall". That is a fact about the model in general; it says nothing about your hardware, your quantisation, or how it behaves on your rig. The breakdown is now kept, and the coding and everyday-chat picks come from how the models actually answered those questions here, marked so you can tell which is which. A pick needs at least three questions behind it and a clear margin over the runner-up, so a lucky answer is never presented as a verdict — which does mean the shorter ten-question run is too thin for most of them, and twenty is the first length that tells you much.
• RigMatch Chat uses the same measurements. When it summarises a long conversation to free up room, it picks the model that actually followed instructions best on your machine, which is often not the one with the highest headline score — that score rewards speed, and the fastest model is usually the smallest.
• Under the hood: the last outstanding security advisory in the app's build tools is cleared, and a test now fails the build if any status colour drops below the readable threshold again.

Downloads

• Windows: RigMatch-*-win-x64.exe
• macOS Apple Silicon: RigMatch-*-mac-arm64.dmg
• macOS Intel: RigMatch-*-mac-x64.dmg
• Linux x64: RigMatch-*-linux-x86_64.AppImage or RigMatch-*-linux-amd64.deb
• Linux ARM64 / Jetson: RigMatch-*-linux-arm64.AppImage or RigMatch-*-linux-arm64.deb

Already running an older install? It updates itself to this version automatically — nothing to re-download.

macOS first launch

RigMatch for macOS is currently an unsigned beta distributed outside the App Store. On first launch, macOS may say the developer cannot be verified or that the app was downloaded from the internet.

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

After that first approval, RigMatch opens normally by double-clicking. If macOS says the app is damaged after copying it to Applications, run this Terminal command once:

xattr -cr /Applications/RigMatch.app

Apple referenc...

v0.4.3-beta — Macs Can Open It Now

What's new in 0.4.3 — Macs Can Open It Now

• RigMatch now opens on Apple Silicon Macs. Every previous build failed with "RigMatch is damaged and can't be opened" — and unlike the usual unsigned-app warning, there was no "Open Anyway" button in Privacy & Security to get past it, so there was genuinely nothing you could do. Apple Silicon refuses to run an app whose signature does not match its contents, and the build was rearranging the app after macOS had signed it and never signing it again. It now signs itself at the end of every build, and the build fails outright if that does not work rather than shipping something that cannot start.
• If you hit this on an older build and gave up: sorry. It was not your Mac. Download this version and it should open normally — you will still see the ordinary "unidentified developer" prompt on first launch, which the download page explains how to clear.
• The instructions for repairing an older install were also wrong. They only cleared the download flag, which was never the problem on Apple Silicon. The correct commands are now on the download page.
• Intel Macs, Windows, and Linux were not affected by this and need no action.

Downloads

• Windows: RigMatch-*-win-x64.exe
• macOS Apple Silicon: RigMatch-*-mac-arm64.dmg
• macOS Intel: RigMatch-*-mac-x64.dmg
• Linux x64: RigMatch-*-linux-x86_64.AppImage or RigMatch-*-linux-amd64.deb
• Linux ARM64 / Jetson: RigMatch-*-linux-arm64.AppImage or RigMatch-*-linux-arm64.deb

Already running an older install? It updates itself to this version automatically — nothing to re-download.

macOS first launch

RigMatch for macOS is currently an unsigned beta distributed outside the App Store. On first launch, macOS may say the developer cannot be verified or that the app was downloaded from the internet.

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

After that first approval, RigMatch opens normally by double-clicking. If macOS says the app is damaged after copying it to Applications, run this Terminal command once:

xattr -cr /Applications/RigMatch.app

Apple reference: https://support.apple.com/guide/mac-help/open-a-mac-app-from-an-unknown-developer-mh40616/mac

What's Changed

• Ad-hoc sign the macOS bundles so Apple Silicon will run them by @DaveEuson in #10
• Release 0.4.3 — Macs Can Open It Now by @DaveEuson in #11

Full Changelog: v0.4.2-beta...v0.4.3-beta