README / README.md
MastLLM's picture
Links point to the Mast-LLM organization
479943d verified
|
Raw History Blame Contribute Delete
2.59 kB
metadata
title: MastLLM
emoji: 🤖
colorFrom: green
colorTo: gray
sdk: static
pinned: true
MastLLM: other labs keep building bigger, we build small

MastLLM

Other labs keep building bigger. We build small. MastLLM makes compact, open AI models that run on the hardware you already own, and the apps that put them to work. The Mast family of sub-1B models is coming soon.

openTARS

Our flagship product. openTARS is an open source AI assistant that runs on your own computer. Ask in your own words: it opens programs, clicks buttons by their name, types, searches the web, runs commands and looks at your screen. By default nothing leaves your machine, and there is no account or subscription. It runs on Linux, Windows and macOS.

Models

Model What it is Size
Gargantua 2.5 The main model: Gemma 4 E4B fine-tuned for the assistant, with vision and audio 5.3 GB + 0.99 GB projector
Gargantua Lite Gen 3 Gemma 4 E2B, for computers without a graphics card, with vision and audio 3.4 GB + 0.99 GB projector
Gargantua Pro Gen 1 The largest model: Gemma 4 12B with vision. Preview, training not finished 7.4 GB
Gargantua Code Gen 2 Writes, fixes and runs code: Qwen2.5-Coder 7B fine-tuned for the tools 4.7 GB
Gargantua Pi Gen 1 For small boards such as the Raspberry Pi, with four tools 300 MB
Murph A tiny request classifier that picks the kind of task in 0.2 ms, with no GPU 4 MB

The openTARS models are published as GGUF files with SHA-256 checksums. Every model card lists the tests we ran and the known limits. These are our own tests on one machine, not public benchmarks.

The app

  • openTARS on Hugging Face: a mirror of the app and the Ubuntu/Debian package.
  • GitHub: source code, releases and issues (MIT).
  • Website: downloads and install commands for Linux, Windows and macOS.

Contact

projectopentars@gmail.com

The openTARS code is MIT. Each model keeps the license of its base: Gemma 4 (Apache-2.0), Qwen2.5-Coder (Apache-2.0) and FunctionGemma (Gemma terms).