SPECTRA framework introduced to compress LLMs by three times

Developers have introduced SPECTRA, a new framework for compressing large language models that reduces their size by 3x with virtually no loss in accuracy. The technology is focused on optimizing model performance on devices with limited hardware resources.

A key feature of the method is that it does not require data calibration. Thanks to this solution, standard 7-billion-parameter models can be run on hardware with 4 GB of VRAM. Furthermore, the compression process and the algorithm itself require less than 50 MB of RAM, making it one of the most lightweight tools in its class.

The developers note that this approach significantly lowers the barrier to entry for running modern LLMs locally. This opens up possibilities for integrating powerful language models into consumer electronics and mobile systems, where memory capacity is a critical constraint. Detailed technical specifications and test results are available in the team's official research report.