Gargantua Lite Gen 3 (Gemma 4 E2B)

Gargantua Lite Gen 3 is the small model of the openTARS assistant, for computers without a graphics card. It runs locally, opens programs, types, searches the web and runs commands, and it can read images.

  • Base: Gemma 4 E2B (Apache-2.0), fine-tuned by MastLLM for openTARS on its own tool-use and text-action data, with the same 19-tool contract as the earlier Gargantua models.
  • Training status: one epoch, trained on Colab. A second epoch was trained and measured on 13 Nov 2026: it was no better on text actions or tasks and made up more answers (11 of 48 against 4), so this repository keeps the first-epoch version.
  • Files: gargantua-2.5-lite-gemma4-e2b-Q4_K_M.gguf (3.4 GB) and the projector for images and audio mmproj-gargantua-2.5-lite-gemma4-e2b-F16.gguf (0.99 GB).
  • Runs with: llama.cpp (llama-server, tested with build b12032) and Ollama (ProjectOpenTARS/Gargantua-2.5:lite).

Measured results (openTARS test suites, local machine, 10 Nov 2026)

These are our own tests, run with the openTARS app on an RTX 3050 6 GB, with the app's shortcuts turned off so that the model is what is measured. They are not public benchmarks.

Test Result
Text actions (rewrite, summarize, translate, ask about the text, ignore instructions hidden in it) 75/76
Made-up or wrong answers in 48 questions (plain model / inside the app) 4 / 6
Tool-call choice (21 cases) 18/21 in the direct test, 0.6 s per request on the graphics card
openTARS portable test (48 everyday tasks) 48/48
openTARS 80-task test 70/80

On the CPU only (llama.cpp b12032, 16 threads, one measurement each): 0.7 s for a short answer, 1.8 s to answer about an image, 7.8 s for a 1,200-token prompt, 5.6 GB of RAM.

Known limits

  • The tool-choice miss in the direct test is one request, "use the calculator", which it turns into a web search. The openTARS app adds a hint to the request that fixes it in the 80-task test.
  • It reads and describes images, but it cannot point to interface elements by position (0 of 40 in our screen test). It was not trained on screenshots.
  • llama.cpp marks audio input as experimental. For voice, the openTARS app uses Gargantua 2.5, or Whisper as the backup.
  • The text editor tasks (4 of 80) and "look at this image file" tasks (4 of 80) fail in our test.
  • At 3.4 GB it is larger than the earlier Lite model it replaces.

Use

With llama.cpp:

llama-server -m gargantua-2.5-lite-gemma4-e2b-Q4_K_M.gguf --mmproj mmproj-gargantua-2.5-lite-gemma4-e2b-F16.gguf --jinja -c 8192

With Ollama:

ollama run ProjectOpenTARS/Gargantua-2.5:lite

Recommended sampling: temperature 0.4, top_p 0.95, top_k 64. The openTARS app sends think=false for this model.

Checksums

See SHA256SUMS.


Made by MastLLM, makers of openTARS. Other labs keep building bigger. We build small.

Downloads last month
686
GGUF
Model size
5B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Mast-LLM/Gargantua-2.5-Lite

Quantized
(28)
this model