How to use from
llama.cpp
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf OpenTARS/Gargantua-2.5:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf OpenTARS/Gargantua-2.5:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf OpenTARS/Gargantua-2.5:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf OpenTARS/Gargantua-2.5:Q4_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf OpenTARS/Gargantua-2.5:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf OpenTARS/Gargantua-2.5:Q4_K_M
Use Docker
docker model run hf.co/OpenTARS/Gargantua-2.5:Q4_K_M
Quick Links

Gargantua 2.5 (Gemma 4 E4B)

Gargantua 2.5 is the main model of the openTARS assistant. It runs locally, opens programs, types, searches the web and runs commands, and it can read images.

  • Base: Gemma 4 E4B (Apache-2.0), fine-tuned by the openTARS project on its own tool-use and text-action data, with the same 19-tool contract as the earlier Gargantua models.
  • Training status: the first training run finished: one epoch, 1,949 steps. Steps 1 to 1,200 ran on a Colab T4 (fp16); the last 749 ran on an RTX 3050 6 GB (bf16, with the decoder layers in 4 bits and the 1.8% longest examples, above 2,048 tokens, left out so it fits in 6 GB).
  • Files: gargantua-2.5-gemma4-e4b-Q4_K_M.gguf (4.9 GB) and the projector for images and audio mmproj-gargantua-2.5-gemma4-e4b-F16.gguf (0.99 GB).
  • Runs with: llama.cpp (llama-server, tested with build b12032) and Ollama (ProjectOpenTARS/Gargantua-2.5:latest).

Measured results (openTARS test suites, local machine, 13 Nov 2026)

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

Test Result
Tool-call choice (21 cases) 21/21, 1.2 s per request on the graphics card, 3.1 s on the CPU
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) 6 / 7
openTARS portable test (48 everyday tasks) 47/48
openTARS 80-task test 68/80
Speech to text, 6 short spoken requests (3 English, 3 Portuguese, synthetic voices), CPU only 6/6, 2 to 3.5 s each

On the CPU only (llama.cpp b12032, 16 threads, measured on an earlier export of the same model): 1.4 s for a short answer, 3.9 s to answer about an image, 14.1 s for a 1,200-token prompt, 8.6 GB of RAM.

Known limits

  • It reads and describes images, but it is weak at pointing to small interface elements by position (0 of 40 in our screen test). It was not trained on screenshots; the openTARS app clicks by accessibility names instead.
  • Speech was only tested on short synthetic phrases; real voices, noise and long recordings are not measured yet. llama.cpp marks audio input as experimental.
  • The text editor tasks and "look at this image file" tasks fail sometimes in our 80-task test.
  • It can still get facts wrong. The openTARS app adds a warning and web checks for those cases.

Use

With llama.cpp:

llama-server -m gargantua-2.5-gemma4-e4b-Q4_K_M.gguf --mmproj mmproj-gargantua-2.5-gemma4-e4b-F16.gguf --jinja -c 8192

With Ollama:

ollama run ProjectOpenTARS/Gargantua-2.5:latest

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

Checksums

See SHA256SUMS.

Downloads last month
169
GGUF
Model size
7B 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 OpenTARS/Gargantua-2.5

Quantized
(21)
this model