Translated with AI from the Russian original.
The development could reduce the cost of running AI services and lower hardware requirements
The MWS AI team has opened access to the source code of CoSpaDi, a method for compressing large language models without fine-tuning. The development allows reducing the size of neural networks while maintaining the accuracy of their responses. The code is published on GitHub.

The CoSpaDi (Compressing LLMs via Calibration-Guided Sparse Dictionary Learning) method was developed by researchers at the MWS AI Fundamental Research Center. It is based on a more compact representation of neural network weights: model parameters are encoded using combinations of a small number of elements from a specially formed dictionary.
During testing on Llama and Qwen models with 0.6–14 billion parameters, the method achieved a size reduction of 20–40%. At the same time, response accuracy was on average higher than that of SVD-LLM, one of the common approaches to neural network compression.
The technology could be in demand by companies that run language models on hardware with limited memory. It will help reduce infrastructure costs, deploy models within the corporate perimeter, or simultaneously serve more users on existing capacities. In addition, CoSpaDi can speed up the performance of neural networks—the result depends on hardware and software implementation.











