---
license: apache-2.0
base_model: Qwen/Qwen3.5-2B
library_name: llama.cpp
tags:
  - gguf
  - qwen3.5
  - vision
  - tool-use
  - local
language:
  - pt
  - en
---

# Gargantua 2 (2B)

Gargantua 2 is the main model of the [openTARS](https://project-opentars.github.io/OpenTARS/) assistant. It runs locally, opens programs, types, clicks, searches the web and runs commands, and it can look at a screenshot.

- **Base:** Qwen3.5 2B (Apache-2.0), fine-tuned by the openTARS project on its own tool-use, text-action and screen data, including code examples.
- **Files:** `gargantua-1.2-qwen3.5-2b-Q4_0.gguf` (1.2 GB) and the vision projector `mmproj-gargantua-qwen3.5-2b-Q8_0.gguf` (0.36 GB).
- **Runs with:** llama.cpp (`llama-server`), and with Ollama (`ProjectOpenTARS/Gargantua-2:latest`).

## Measured results (openTARS test suites, local machine)

These are our own tests, run with the openTARS app on an RTX 3050 6 GB. They are not public benchmarks.

| Test | Result |
|---|---|
| 80-task Lab suite (tools, text, vision, refusals) | 73/80 |
| Portable test, 48 tasks (`opentars --prova`) | 46/48 |
| Text actions (rewrite, summarize, translate, ask about text) | 71/76 |
| Tool-call choice (21 cases) | 21/21 |
| 100-task suite with side-effect check (`--cem`, 1 Nov 2026) | 94/100 |

Known limits: it can still invent facts about things it does not know (the app adds a warning and web checks for those cases), it is weak at reading small text in screenshots, and it can answer a greeting with a web search.

## Use

With llama.cpp:

```bash
llama-server -m gargantua-1.2-qwen3.5-2b-Q4_0.gguf --mmproj mmproj-gargantua-qwen3.5-2b-Q8_0.gguf -c 8192
```

With Ollama:

```bash
ollama run ProjectOpenTARS/Gargantua-2:latest
```

Recommended sampling: temperature 0.4, top_p 0.8, top_k 20, repeat_penalty 1.05. The openTARS app sends `think=false` for this model.

## Checksums

See `SHA256SUMS`. The GGUF file's SHA-256 matches the layer already published on Ollama (`99b7dd09…`).