openTARS

Changelog

What changed in each version of openTARS. Updates install from the same package: your chats and settings are kept.

openTARS 5.3 Tesseract, update of November 22

  • openTARS is now made by MastLLM. Same app, same team, new home: the code lives at github.com/MastLLM/OpenTARS, the models at huggingface.co/Mast-LLM, and the site at mastllm.github.io/OpenTARS. Updates and model downloads now go straight to the new addresses; the old links keep redirecting.

openTARS 5.3 Tesseract, update of November 15

  • Native apps for Windows and macOS. openTARS now ships as a ready-to-run app: opentars-setup.exe on Windows and a signed .dmg on macOS, for Apple Silicon and Intel. Python, Qt and the llama.cpp engine are inside, so installing no longer downloads and builds anything.
  • One-click updates that just work. Windows and macOS update from the latest release, verified by SHA-256, and reopen by themselves. After this update, Windows and macOS walk through the welcome setup once more.
  • Windows engine fix (November 15, later). The first Windows build of this update could stop with "llama-common.dll was not found" when the AI engine started. The engine is now packaged whole, and the build checks that both engine variants (processor and Vulkan) actually start before it is published. The installer also carries proper version and publisher details, and each build is scanned with Microsoft Defender before release.
  • A model store on Hugging Face. Hand-picked GGUF models (Qwen3.5, Gemma 4, Ministral 3, Llama 3.2, Phi-4 mini, gpt-oss and more) with fixed checksums, plus search across the Hugging Face community.
  • All in on llama.cpp. The engine sizes each model to your GPU memory, falls back to the processor if the graphics driver fails, restarts itself if it stops, and replaces an old engine after an update.

openTARS 5.3 Tesseract

November 13, 2026

  • Rebuilt on llama.cpp: one engine for every model, downloaded by the app the first time it opens, with no password and no separate AI program. Coming from 5.2, your Gargantua models are copied over from Ollama. Ollama stays available in the developer settings.
  • Gargantua 2.5, on Gemma 4: the new main AI picked the right tool in 21 of 21 requests and handled 75 of 76 text actions in our tests. Gargantua Lite Gen 3 (Gemma 4 E2B) comes along for lighter machines.
  • Voice, understood directly: with Gargantua 2.5, what you say goes straight to the model, in English or Portuguese. Whisper stays as the backup.
  • Missions in a virtual machine: on Linux with KVM, Gargantua Pro works inside a disposable computer. A new window in Probes shows its screen live, the steps and every site it visited, and you choose which files come back.
  • Model store with search: filter the list as you type, or search the whole Ollama library from Settings. Gargantua Pro is in the store.
  • New terminal look: a framed prompt, a status bar with the model, the engine, the folder and the branch, formatted answers, and the logo when the chat opens.
  • One-click updates on every system: Windows, macOS and other Linux now update from the Updates page instead of copying a command.
  • A stronger assistant: calendar (.ics), notes, e-mail, music controls and system settings by sentence on Linux, Windows and macOS. Everything new is off until you turn it on, and e-mail never sends without your yes.
  • Clicks in Chromium and Electron apps: with your permission, openTARS reopens an app with accessibility on so it can click its buttons by name.
  • Fixes: phrases like "the game freezes and the screen goes black" no longer lock the screen; "save my file" no longer turns into a memory; e-mail checks the server certificate; a newer model with the same file name now replaces the old one; the installers no longer install Ollama unless you ask.
  • Verify your download: a new page explains the checksum and the GitHub attestation of each release.

openTARS 5.2 Ranger, update of November 10

November 10, 2026

  • Works with the current Ollama again: with the window open, every request failed with “time: missing unit in duration”. The app now asks Ollama to keep the model loaded in the form it accepts. If you saw that error, update and reopen openTARS.
  • No more endless update offers: after you updated from inside the app, it kept offering the same package again. It now remembers the package by both of its identifiers, so one update is enough.

openTARS Pi 1.1

November 7, 2026

  • It finds what you mean: “open the report” opens Documents/relatorio-2026.txt. Wrong case, a missing extension or a typo get fixed before anything runs, and it tells you which file it used. Deleting never guesses.
  • Tasks with several steps: “create a folder backup and copy notas.txt into it”, “find WARN in the log and save those lines to a file”.
  • Answers in words: counts, sizes, the biggest file, “is the SD card full?”, “is there room for 2 GB?”.
  • Better answers about the world: it looks for the sentence with the kind of answer you asked for, a name, a year, a height or a place.
  • A bigger brain when you want one: use the Gargantua on your computer, over your own network.
  • Updates in place: opentars update downloads only what changed and can undo.
  • Capability test: 30 of 40 everyday tasks it had never seen, up from 19 of 40 in 1.0.

openTARS Pi 1.0

November 6, 2026

  • New edition for small boards: a terminal assistant for Raspberry Pi, Orange Pi and other 64-bit boards, with its own model, Gargantua Pi (270M, 300 MB). It installs with one line and no sudo, and runs on llama.cpp.
  • Pins by name: read and switch GPIO in plain words, with your yes on the board for every change.
  • A service for the house: optional, off by default, with a token. Requests from the network are approved on the board.

5.2 Ranger

October 24, 2026 · updated November 9

  • Gargantua 2 Lite retrained (November 9): ProjectOpenTARS/Gargantua-2:lite was trained again. In our tests it handles text actions 67 of 76 times (before: 60), invents fewer facts when run inside openTARS (9 errors in 48 questions, before 17), and got 96 of 100 everyday tasks right on the processor alone (before: 93). Run ollama pull ProjectOpenTARS/Gargantua-2:lite to get it.
  • Gargantua 2 retrained (November 9): ProjectOpenTARS/Gargantua-2:latest (the 2B) was trained again and handles text actions 76 of 76 times in our test (before: 72 of 76), says it does not know about people and things that do not exist 16 of 16 times inside the app, and corrects a false premise more often. Run ollama pull ProjectOpenTARS/Gargantua-2 to get it. The new files are also on Hugging Face.
  • Fix (November 4): a local Gargantua with a slash in its name is treated as Gargantua again, so it no longer turns on reasoning and comes back with no text. The model stays loaded while the window is open and returns to the normal timeout in the tray. A greeting no longer triggers the web check.
  • Fix (November 2): “Show in folder” in the file search now opens the file manager with the file selected, instead of only opening the folder.
  • Fix (November 1): the portable test (opentars --prova) runs without the desktop: before, it could open apps and type into the open windows of the machine where it ran. Use it only after this update.
  • Patch (October 27): the web check that confirms a fact only runs for a real question: a request for a poem or a text no longer triggers a web search. The same search does not run twice in one request, and each request stops after three searches.
  • Patch (October 26): updating works with the assistant on (openTARS restarts by itself after installing); the voice download shows a progress bar; “More models” marks the models that fit your graphics card and memory; web search through SearXNG is on by default (without it, the search opens in the browser); less background work while idle.
  • Make images (October 25): ask “draw…” or “generate an image of…” in the chat. openTARS uses Bonsai Image 4B by PrismML, an optional install (about 9.5 GB, an NVIDIA card with 6 GB or a Mac with Apple Silicon) from Settings → More models, with no administrator password. It runs in its own process with no open port, takes the other models off the card first and tells you when the card has no room (a game using it, for example). On our RTX 3050 the first 512-pixel image took about a minute while the model loaded, the next ones about 5 s. Saved in Pictures/openTARS.
  • Small photo models (October 25): remove the background, upscale 4x, select by clicking and remove an object, all on the processor and downloaded with openTARS (about 300 MB from Hugging Face, each file checked against a fixed SHA-256). Licenses: MIT and BSD.
  • Edit photos from the chat (October 25): attach or name a photo and say what to do (“make it square and brighter”, “remove the background”); the copy appears in the answer.
  • Gargantua 2 Lite always installed (October 25): it is the only AI the assistant uses in the background (text actions, Probes) and it unloads when idle; text actions use Gargantua 2 only for long texts. More models now starts with the openTARS models, the Lite first.
  • Photo editing in the assistant (October 25): “Edit a photo” is now in the launcher and in the icon near the clock: choose the photos and the panel offers crop, brighten, remove the background, upscale and remove an object.
  • Fixes (October 25): the chat no longer scrolls past the last message; on a Mac the opentars command is on the Terminal's PATH; on Windows the terminal shows colors and a clean prompt; the Windows shortcuts and taskbar show the openTARS icon.
  • Windows and macOS: one-command installers with no administrator password (Python through uv, Ollama and openTARS in your user folder). On Windows, openTARS opens and closes programs by name (Start menu apps included), moves windows, and clicks buttons and fills fields by their name through UI Automation; Gargantua's file commands run in Python, since Windows has no bash. On macOS, openTARS.app is built and signed on your Mac, and clicking by name uses Accessibility. Both pass our automated tests on GitHub's Windows and macOS machines. Windows had its first run on a real laptop on October 24: the first-run window froze while checking the graphics card and step 4 did not continue; both are fixed. No real Mac has run openTARS yet, and voice and the global shortcuts have never run outside Linux.
  • Other Linux: a one-line installer for Fedora, Arch and openSUSE (packages from dnf, pacman or zypper, the rest in your user folder), tested in containers, including the real window on a virtual screen.
  • The launcher: the quick bar with nothing typed offers five things: do a task on the computer, ask about the screen, find a file, search the web and create a reminder. Finding a file looks at file names, folders and the text inside documents (PDF too), with no index and no model.
  • Text actions: select text in any app and press Ctrl+Alt+R (Control+Option+R on a Mac): fix, more formal, simpler, summarize, translate between English and Portuguese, or ask about the text. Replace puts the result back and restores what you had copied. Text in, text out: no tools run, and the selected text is treated as material, never as instructions. In our test of 76 text tasks Gargantua 2 got 71 right and the Lite 64. The shortcut comes on by default; turn it off or change it in Settings → Assistant.
  • Actions on selected files: summarize, rename by content with a preview you can edit, move to a folder, or send to the trash after you confirm; renaming and moving can be undone while the panel is open.
  • Edit a photo with a sentence: “crop it square and make it brighter”, in English or Portuguese: crop (square, 16:9, 4:3, 3:2), rotate, mirror, brighter or darker, contrast, black and white, shrink for email and change the format. A before and after preview first, and the result is always a new copy next to the original, without location and device data. No image model: a 12-megapixel photo takes 0.35 s.
  • Assistant: an icon next to the clock with the text actions; text actions use Gargantua 2 Lite, which leaves the memory by itself 3 minutes after the last use (when idle, the AI takes no memory or graphics card), openTARS starts with the system in that icon, and it can suggest help based on the program you are using. All on by default, from the first run; anything you turn off in Settings → Assistant stays off.
  • When it is unsure: on a factual question, openTARS reads how confident the model was in its answer. With web search on, it checks with a search; without it, it adds a note under the answer. It is a hint, not a detector: in our check on questions it had not seen, it flagged 2 of 5 wrong answers from Gargantua 2 and 4 of 6 from the Lite, and 3 right answers from each.
  • Gargantua 2 and Lite retrained with the 5.2 examples (published on October 23): saying it does not know, everyday conversation, text actions, answering from search results with the source, and Windows and macOS. 73 of 80 everyday tasks for Gargantua 2 and 68 for the Lite. openTARS now tells you when the Gargantua you have is older than the published one, with a button to update it.
  • Check it yourself: opentars --prova runs 48 everyday tasks with your model in a temporary folder and saves a report (Gargantua 2 does 46 of 48 on Linux); opentars --diagnostico saves a report about your system; opentars --desinstalar removes openTARS and asks before deleting your data or models.

5.1 Tesseract

October 16, 2026 · updated October 23

  • Gargantua 2 and Lite retrained (October 23): the same ProjectOpenTARS/Gargantua-2, trained again with everything from before plus examples of saying it does not know, everyday conversation, working on text you select and answering from search results with the source. In our test of 80 everyday tasks it did 73 right (the text editor went from 2 to 4 of 4) and ran no dangerous command. Asked about people and things that do not exist, it said it did not know 16 of 16 times; it still corrects a false premise only 2 of 6 times. Run ollama pull ProjectOpenTARS/Gargantua-2 to get it. Gargantua Lite was retrained the same way: 68 of 80 tasks (66 before), it says it does not know 15 of 16 times (0 before) and finds a button 9 of 15 times (4 before); it sometimes uses a web search where the calculator was asked for. Run ollama pull ProjectOpenTARS/Gargantua-2:lite.
  • More models and Manage models (October 21): two new pages in Settings (Base and Pro modes). More models has 58 models you can add with one click: Bonsai 2, image generation (FLUX.2 klein and Z-Image Turbo, which Ollama runs on macOS first, so they may download here and not run), general use, vision, code and a group of large models that need a big graphics card. Downloads go through Ollama, with a progress bar, a Cancel button and a check that the disk has room. Manage models lists what Ollama has installed, largest first, with its size, and lets you update a model or delete it (it asks first, and warns for the official openTARS models). Models that only draw images are never picked to chat.
  • Images in answers (October 21): when an answer includes an image from your own computer (![text](/path/photo.png)), the window now shows it, scaled to fit. Only local image files are shown: addresses from the internet are never loaded, they stay plain links.
  • Gargantua 2 is the main AI, and Cooper is discontinued (October 18): ProjectOpenTARS/Gargantua-2 is a Qwen3.5 2B trained by us, 1.6 GB (the first Gargantua was 2.5 GB), and it sees the screen itself. In our test of 84 everyday tasks it did 72 right and ran no dangerous command. The installer and the first-run window download it, Settings offers it to anyone who has the old Gargantua (nothing is removed), and it is preferred when both are installed. Cooper, the separate vision model, is discontinued: it is no longer downloaded, trained or supported, and opentars --cooper only says so. openTARS asks Gargantua 2 the same screen questions, with the model that is already in memory, so there is no extra download and no model swap. An old Cooper still installed is used only when there is no Gargantua 2. Older openTARS versions keep working with the old Gargantua.
  • Gargantua reads web searches (October 18): with web search on (SearXNG, Settings → Cloud), a search only asks SearXNG and Gargantua answers from the first results instead of from memory. Nothing opens on your computer unless you ask for it (“open it in the browser”) or search a site with its own search, like YouTube. If SearXNG is down, it says so instead of opening Google. Gargantua does this through the search tool it was trained with, so nothing changes in its training.
  • Probes in your own words work with Gargantua 2 (October 18): the models were never trained to turn a mission into a plan (the Lite got 2 of 16 plans right). openTARS now builds the plan itself when the mission is clear (a folder, a page, a feed, the battery, the temperature, a program closing or using too much memory, a git project) and only asks the model otherwise, forcing a valid format and fixing targets that do not exist. In our tests on 16 missions it got 15 of 16 for Gargantua 2 and 14 of 16 for the Lite; missions that ask to delete, move, send or install are refused by the code.
  • Answers in your language, and honest about the browser (October 18): Gargantua answered in Portuguese to about 4 in 10 requests written in English; openTARS now tells it to answer in English when you write in English (0 in 32 in our test, and the English tasks of our proof did not get worse). When a web search only asked SearXNG and the answer still says it “opened the results in the browser”, openTARS adds a note that nothing was opened.
  • Better tool use (October 18): a short follow-up like “use the tool”, “do it for real” or “try again” now carries the previous request, so the AI does that request instead of guessing; a site name asked as an app (“open google”) opens the site instead of a browser that only has the word in its name; after a web search the AI is told it has not read the results, and if it still answers from memory, openTARS says so under the answer. Murph also learned more: in our evaluation it went from 163 to 170 of 175 phrases in five languages.
  • Murph and Probes (October 18): Murph now recognizes short chat messages that only continue a conversation, so they are not taken for commands; a Probe mission that asks to delete, install or change anything is refused by the code itself, and every Probe gets a name taken from its mission.
  • New Gargantua Lite (October 17): 0.6 GB with vision, down from 1.1 GB without it, and 70 of 84 everyday tasks in our test, up from 63. It runs on the processor alone. Computers without a graphics card and with little memory download ProjectOpenTARS/Gargantua-2:lite. (October 18) The Lite was retrained with screens: it now finds the right button in 29 of 40 tries when openTARS asks the way it does to click, up from 12, and does 68 of 84 everyday tasks (the previous one did 69 in the same run). Gargantua 2 still does it better (34 of 40) and Settings offers it to Lite users. If you downloaded the Lite on October 17 or 18, run ollama pull ProjectOpenTARS/Gargantua-2:lite to get the new one.
  • Fixes (October 17): openTARS no longer pushes the AI to act when it only guessed that a chat message was a request; a model that needs a newer Ollama now says how to update it.
  • Three modes: Lite (chat, microphone, attachments), Base (adds history, memory, routines, Probes and the cloud) and Pro (adds advanced settings and diagnostics). Chosen on first run, switchable in Settings; updating from 5.0 keeps you in Pro.
  • Probes: small watchers for a folder, a web page, news and releases (RSS, a YouTube channel, GitHub releases), this computer (battery, temperature, CPU, internet, system updates), a program or a git project. Creating one is a three-step guide with a fixed rule and no AI; "in my own words" still turns a mission into a plan you check. They only read; pages and feeds only from public addresses, at most every 15 minutes. Off by default; keys, passwords and browser data are off limits. Optional actions you choose: move what it finds to a folder, or send it to the trash after you approve each time.
  • Cooper 1.2: the installer and the first-run window now download ProjectOpenTARS/cooper:1.2 (SmolVLM2 500M, about 636 MB) instead of the older 2.7 GB model. An older Cooper that is already installed keeps working; when both are installed, 1.2 is used.
  • Fixes: the window no longer crashes on computers with a damaged font cache (openTARS keeps its own); full-screen terminal programs such as htop or vim open in a terminal window instead of failing; with both ~/Documents and ~/Documentos, a file is found in either one.
  • Alerts: a bell next to the clock shows how many alerts you have and opens a panel on click, even with the window closed; inside the app, the same bell at the top and a bar for new alerts.
  • Gargantua Lite on computers without a graphics card and with little memory, from the first run and the installer.
  • Terminal: opentars -p for scripts (with JSON output), resume, compact, context, plan and auto modes, file edits with a diff and your approval, undo, AGENTS.md, /init and !command.
  • Cloud with your Ollama account, no key to copy. The footer now also warns when the chosen model is an Ollama cloud model.
  • Web search through SearXNG, off by default.
  • When a model fails mid-request, the automatic switch never drops to a much smaller model.

5.0 Tesseract

October 6 to 15, 2026

October 15 update: two small Gargantuas are now available as optional downloads, Gargantua Code and Gargantua Lite (see the models section on the home page).

October 15 update: Gargantua now uses its tools much more reliably. openTARS was asking Ollama to turn reasoning off in a way that made Gargantua think it should answer right away, so sometimes it replied without doing the task. In our tests Gargantua used its tools in 18 of 18 tries, up from 15.

October 13 update: closing openTARS now unloads only the models it loaded itself (and Cooper), instead of every model in Ollama, so other programs that share Ollama keep theirs. The “unload all models” command still unloads everything. Also removed a voice help line that mentioned options that do not exist.

October 11 update: the mascot is now called KIPP, after the third robot of Interstellar, and its tips were rewritten. The Update button now really installs the new package (before, it could say it had updated without changing anything), checks the result, and looks for updates every 6 hours while the window is open.

A new window built in Qt, on top of the lean 4.6 core: Ollama underneath, Cooper looking at the screen and Gargantua as the official AI.

  • First run in five steps: language, Ollama, optional cloud, downloading Gargantua with a progress bar you can pause, and a quick test.
  • History by day with search, a model picker, a Stop button and the steps of each task with their timings.
  • Risky commands turn into a card with Allow and Deny.
  • Voice is back, fully local: hold to talk, continuous, or the wake word “TARS”.
  • Attachments, a visible memory you can edit, and scheduled routines that never decide alone.
  • An optional cloud with your own account, off by default and never used by the automatic model choice.
  • Updates noticed by the app itself, verified against GitHub and installed after you confirm.
  • A quick bar on Ctrl+Alt+Space.
  • A code audit with about 50 fixes: atomic settings files, more commands that ask first (rm with wildcards, find -exec rm, curl | sh, ssh), and the cloud key never passed to commands.

Updated on October 10

  • A mascot that walks along the chat, shows what the AI is doing and naps after a few quiet minutes.
  • A terminal styled like today's AI command-line tools: welcome panel, steps you can follow, a spinner and a clearer /help.
  • Fixes: opentars --voz, turning the update check fully off, commands ending in &, and more mv commands that ask first.

4.6 Endurance

October 4 and 5, 2026

The lighter release: it only removed what was heavy or unused.

  • Ollama only: external servers and importing .gguf files from other programs were removed.
  • Only Cooper looks at the screen: OCR, the zoom grid and the fallback vision model were removed.
  • Less dead code and fewer duplicated command-line options.
  • Voice, attachments and routines were removed here and came back in 5.0.

Still available for download at the end of the install section.

4.5 Endurance

October 3 and 4, 2026

  • Cooper, a vision model fine-tuned to find buttons, read fields and list windows on a screenshot.
  • WSL 1 and 2, with the Ollama running on Windows found automatically.
  • A terminal mode for servers and SSH sessions without a screen.
  • From 28 to 19 tools, and more than 40 fixes, including Stop while a model loads and a real right click.

4.0 Endurance

September 30 to October 2, 2026

  • Gargantua, the official AI: a Qwen3 4B trained to use openTARS, downloaded by the installer.
  • Murph, a tiny classifier that decides the type of each request in about 0.2 ms.
  • Longer tasks, the real names of your personal folders, and a notice when a new version comes out.

What comes next is on the home page.