Project Status:

TurboPenguin

Free Software that runs mid-sized LLMs on GNU/Linux without loading the whole model into memory – aiming for more than 80% less RAM at about 95% of the performance.

#Round 5 

TurboPenguin is a Free Software that enables mid-sized LLMs to operate in unified memory without loading the complete model at once. Building on techniques demonstrated on proprietary Apple and Windows platforms, it brings partial model loading to GNU/Linux on PC-compatible hardware. The project targets a reduction of more than 80% in RAM use while retaining about 95% of performance, thereby lowering hardware and energy requirements. The resulting open and reproducible prototype could broaden access to local AI and lay the groundwork for running capable models on smaller devices.

Team

  • Volker Birk
  • Luca Saiu
  • Neal Walfield
  • Marcio Ferreira dos Santos
PROJEKTÜBERSICHT