25 Downloads Updated 21 hours ago
Gargantua 2.5: the main openTARS assistant on Gemma 4 E4B. Uses your computer’s tools, sees the screen and understands audio. Runs locally.
ollama run ProjectOpenTARS/Gargantua-2.5
Applications
ollama launch claude --model ProjectOpenTARS/Gargantua-2.5
ollama launch opencode --model ProjectOpenTARS/Gargantua-2.5
ollama launch hermes --model ProjectOpenTARS/Gargantua-2.5
ollama launch openclaw --model ProjectOpenTARS/Gargantua-2.5
Models
View all →Readme
Gargantua 2.5
Gargantua 2.5 is the main model of openTARS, a local assistant for Linux, Windows and macOS. It chooses which computer tool to use for a request, runs it, and reads the result. It can also look at a screenshot and understand spoken audio. Nothing leaves your computer.
It is built on Gemma 4 E4B and quantized to 4 bits (Q4_K_M). The vision and audio parts come from a projector file included in this package.
Usage
ollama run ProjectOpenTARS/Gargantua-2.5
There is also ProjectOpenTARS/Gargantua-2.5:lite, built on Gemma 4 E2B (3.4 GB), for computers without a graphics card. The openTARS app picks the right one for your machine.
Training
One epoch, 1,949 steps. Steps 1 to 1,200 ran on a T4 GPU in fp16. Steps 1,201 to 1,949 ran on an RTX 3050 with 6 GB, in bf16 with the decoder layers in 4-bit. Examples longer than 2,048 tokens (1.8% of the data) were left out so the run fit in memory.
Results
These numbers come from our own tests, run through the openTARS app with its shortcuts turned off. They were measured on the step-1,200 checkpoint. The final version has not been through the full test yet.
| Test | Result |
|---|---|
| Choosing the right tool (21 requests) | 21 of 21 |
| Text actions: rewrite, summarize, translate, ask about text, ignore instructions hidden in it (76) | 73 of 76 |
| Made-up answers to questions about things that do not exist (48) | 4 |
| Time per tool request, RTX 3050 | about 0.9 s |
| Time per tool request, CPU only, 16 threads | about 3.4 s |
These are not public benchmarks.
Limits
Clicking a button by its position on the screen is weak (2 of 40 in our test); the app looks up elements by their text first. Audio has not been measured on real speech yet. Text editing, counting and chained calculator tasks still fail sometimes. The model can still state things that are false, so the app searches the web or flags low-confidence answers.
License
The base model is under the Gemma 4 terms (Apache-2.0). The openTARS fine-tune and app are MIT.