Gargantua 2: openTARS's local AI for Linux. A small Qwen3.5 model (QAT 4-bit) that uses 19 desktop tools and can see the screen. Runs on CPU.

Details

Updated 5 days ago

5 days ago

ec531c969441 · 1.6GB

model

archqwen35

parameters1.88B

quantizationQ4_0

1.2GB

projector

archclip

parameters331M

quantizationQ8_0

365MB

params

{ "repeat_penalty": 1.05, "temperature": 0.4, "top_k": 20, "top_p": 0.8 }

65B

Gargantua 2

The official AI of openTARS, the 100% local assistant that controls your Linux desktop. It decides what to do, calls the right tool (open an app, press a button by its name, search, create a file, run a command) and looks at the screen when it needs to.

Gargantua 2 replaces Gargantua 1 and Cooper (a separate vision model, discontinued on October 18, 2026): one model now thinks and sees. The old repository, ProjectOpenTARS/Gargantua, stays for older openTARS versions.

Tag	Base	Size	Pull
`latest`	Qwen3.5 2B	1.6 GB (model + vision)	`ollama pull ProjectOpenTARS/Gargantua-2`
`lite`	Qwen3.5 0.8B	0.6 GB (model + vision)	`ollama pull ProjectOpenTARS/Gargantua-2:lite`

Both are quantized to 4 bits (Q4_0) after quantization-aware training, so the model learned to lose little when it shrinks. Needs a recent Ollama (we tested 0.40).

What we measured

Our own tests, on requests the models never saw in training (not public benchmarks), on a graphics card with 6 GB:

	`latest` (2B)	`lite` (0.8B)
Everyday desktop tasks (84, in openTARS’s test)	72 of 84	68 of 84
Dangerous requests refused or confirmed	4 of 4	4 of 4
Picks the right tool (21 requests)	21 of 21	21 of 21
Time per request	0.7 s on a graphics card, about 4 s on a 16-thread CPU	0.3 s on a graphics card
Finds the right button on a screenshot	34 of 40	29 of 40
Reads the text of a field	15 of 15	15 of 15
Checks whether a statement about the screen is true	14 of 15	15 of 15

The screenshot tests use screens drawn by us with the exact answer known, and the question openTARS asks when it clicks. Real screens are messier, so openTARS checks the result of every click.

How it was trained

An open teacher model used openTARS for real, on a virtual screen with a fake home folder, in Portuguese and English.
An automatic checker verified every result: did the calculator show the right number? Does the folder exist? Was nothing deleted?
Only the tasks solved correctly became examples: 2,150 of them, plus 1,500 questions about drawn screens (find a button, read a field, check a statement, describe the windows).
LoRA with quantization-aware training (Unsloth) on a single 6 GB graphics card, 2 epochs, then GGUF Q4_0 with the vision part in Q8_0.

Using it with openTARS

openTARS downloads Gargantua 2 on install (computers without a graphics card and with little memory get lite) and uses it for actions, searches and for looking at the screen. It sends the model the short prompt and the 19 tools it was trained on; with another prompt or tool list it will not work as well.

sudo apt install ./opentars_all.deb

Through Ollama directly, use the chat endpoint with "think": false: these models are Qwen3.5, which thinks by default and may spend the whole answer thinking.

Limits

It makes mistakes in the text editor (2 of 4 in our test), with calculator chains and with counting files.
Looking at an image file (a photo) is weak: it was trained on screens, not photos (1 of 4).
The lite is less precise than the 2B on the screen (29 against 34 of 40) and answers questions with less care. For web search it still needs openTARS, which hands it the results.
It can say something is done when it is not. openTARS checks the tool results and warns, but outside openTARS nothing does.
Portuguese and English only. The 2B sometimes answers in Portuguese to a request in English.

License

The model is a fine-tune of Qwen3.5 (Apache 2.0). The openTARS code is MIT. Contact: projectopentars@gmail.com