---
license: gemma
base_model: google/functiongemma-270m-it
library_name: llama.cpp
tags:
  - gguf
  - gemma3
  - function-calling
  - tool-use
  - tiny
  - local
  - cpu
language:
  - pt
  - en
---

# Gargantua Pi (270M)

Gargantua Pi is the tiniest model of the [openTARS](https://project-opentars.github.io/OpenTARS/) assistant. It is built for small boards and low-power machines: Raspberry Pi, Orange Pi and any computer without a graphics card. At about 300 MB it loads in a moment and answers on the CPU.

It is the model of **[openTARS Pi](https://github.com/project-opentars/OpenTARS-Pi)**, the assistant for small boards:

```bash
curl -fsSL https://github.com/project-opentars/OpenTARS-Pi/releases/latest/download/install.sh | bash
```

It picks between four tools: run a terminal command, search the web, remember a fact, and forget one. It answers in Portuguese and English and politely says so when you ask for something it cannot do, such as looking at the screen.

- **Base:** FunctionGemma 270M, fine-tuned by the openTARS project on its own tool-use data (4 tools: `execute_terminal`, `search_web`, `remember`, `forget`).
- **File:** `gargantua-1.2-pi-functiongemma-270m-Q8_0.gguf` (0.3 GB).
- **Runs with:** llama.cpp and Ollama (the `functiongemma` renderer and parser are built in).
- **License:** Gemma Terms of Use, inherited from the base model. Anyone who redistributes it must pass those terms on.

## Use

With Ollama, create the model from the downloaded file:

```bash
printf 'FROM ./gargantua-1.2-pi-functiongemma-270m-Q8_0.gguf\nRENDERER functiongemma\nPARSER functiongemma\nPARAMETER temperature 0.1\n' > Modelfile
ollama create gargantua-pi -f Modelfile
ollama run gargantua-pi "how much free memory is there?"
```

Pass the four tools in the request so it can call them. The openTARS app does this for you.

## Measured results

Our own tests, not public benchmarks.

- **The model alone:** on 198 command requests it was not trained on, it picks the right program in 150 and calls the terminal in 196. The previous version picked the right program in 6.
- **On a board:** on an Orange Pi Zero 2W (4x Cortex-A53, 1 GB) it generates about 9 tokens per second on the CPU, with the 300 MB file mapped into memory.

At this size it should not be trusted with facts or file paths on its own. openTARS Pi answers those by code and from search, and asks before running anything that changes the board.

## Checksum

SHA-256 of the GGUF: `18e02f9214b821da947fa9b3a2aae96f8679eeba2d7f25646e1032d76dd11775`