Smaller models for existing infrastructure
The pipeline combines established compression techniques with a proprietary quantum-inspired layer intended to recover accuracy lost during compression. The work uses classical computing; quantum-inspired methods do not require a quantum processor.
QuTwo says its evaluation covered seven model families and multiple architectures, including dense and mixture-of-experts models. In one reported example, Qwen2.5-72B was reduced from 145 GB to 14 GB while retaining 90.5% of its knowledge score and 93.5% of its reasoning score.
More control over deployment
The intended benefit is to make capable models easier to run on infrastructure controlled by the customer. The reported percentages describe QuTwo’s own benchmark evaluation, rather than a guarantee of equivalent performance across every application.
