ProjectOpenTARS/ Gargantua:1.2-code

23 1 week ago

ai model optimized for opentars

tools thinking 
ollama run ProjectOpenTARS/Gargantua:1.2-code

Details

2 weeks ago

cc811ac28db1 · 1.1GB

model
qwen3
·
1.72B
·
Q4_K_M
params
{ "repeat_penalty": 1.05, "stop": [ "<|im_start|>", "<|im_end|>" ],
template
{{- if or .System .Tools }}<|im_start|>system {{- if .System }} {{ .System }} {{- end }} {{- if .Too

Readme

Gargantua

The official AI of openTARS, the 100% local assistant that controls your Linux desktop.

Gargantua is a Qwen3 4B fine-tuned for one job: using openTARS tools the right way, on the first try. It opens and closes apps, presses buttons by name, does math in the calculator, creates folders and files, searches the web, writes in the text editor, remembers your preferences and refuses destructive commands without your confirmation.

ollama pull ProjectOpenTARS/Gargantua

How it was trained

Gargantua didn’t learn from hand-written answers. It learned from real openTARS sessions:

  1. An open teacher model (qwen3:8b) used openTARS for real, on a virtual screen with a fake home folder, across 2,560 tasks in Portuguese and English.
  2. An automatic checker verified every result: did the calculator display show the right number? Does the folder exist? Was nothing deleted?
  3. Only the tasks solved correctly on the first try became training examples.
  4. QLoRA (Unsloth) on a single 6 GB GPU, exported to GGUF Q4_K_M.

No outputs from closed models were used. Only open models and verified results.

Why use it

Gargantua A generic model
System prompt + tools ~1,200 tokens ~4,000 tokens
Thinking before acting not needed often needed
Knows the openTARS tools learned from verified examples only reads the instructions
Size 2.5 GB (Q4_K_M) varies

Runs well on a 4–6 GB GPU, and on CPU with 8 GB of RAM (slower).

Using it with openTARS

openTARS 4.0 “Endurance” downloads Gargantua during installation and uses it first for actions and searches in AUTO mode. It automatically sends Gargantua the compact format it was trained on: a short prompt, the 27 tools with one-line descriptions, and no thinking.

  • Pin it manually: /modelo gargantua
  • Check it: opentars --version

Gargantua is tuned for openTARS’s tools and prompt format. It still chats normally on its own, but it shines inside openTARS.

Details

  • Base: Qwen3 4B (Apache 2.0)
  • Method: QLoRA, r=16, 2 epochs, trained on assistant turns only
  • Languages: Portuguese (Brazil) and English
  • Quantization: Q4_K_M
  • License: Apache 2.0

Train your own

The whole training kit is open: one Python file that collects examples, trains on a 6 GB GPU and exports the GGUF. See gargantua/TUTORIAL.md.

🇧🇷 Em português: o Gargantua é a IA oficial do openTARS, um Qwen3 4B treinado com milhares de pedidos resolvidos de verdade no openTARS e conferidos um a um. Abre apps, clica pelo nome, calcula, cria arquivos e pesquisa de primeira, rápido e numa placa de 6 GB. O instalador do openTARS 4.0 já baixa ele.

Follow the project on Instagram: @open.tars