To test how well it worked, the researchers set up a database of about 25 questions on limited topics in Chinese models, including “Who is Winnie the Pooh?” is also included. They tested the response of the modified model against the original Dipsec R1 using OpenAI’s GPT5 as an unbiased judge to rank the degree of censorship in each response. Multiversi says that the parsimonious model was able to provide realistic responses comparable to Western models.
The work is part of a broader effort by Multiverse to develop technology to suppress and manipulate existing AI models. Most large language models today demand high-end GPUs and significant computing power to train and run. However, they are ineffective, says Roman Oris, cofounder and chief scientific officer of Multiverse. They say a compressed model can perform almost as well and save both energy and money.
There is a growing effort in the AI ​​industry to make models smaller and more efficient. Distillation models, such as Dipsec’s own R1-distill variant, attempt to capture the capabilities of larger models by “teaching” them what they know from a smaller model, although they often fall short of the original’s performance on complex reasoning tasks.
Other methods of compressing models include quantization, which reduces the precision of the model’s parameters (boundaries that are set when trained), and pruning, which removes individual weights or entire “neurons”.
“Large-scale AI models are very difficult to build without losing performance,” says Maxwell Venitos, an AI research engineer at Software Informatics. “Most techniques have to compromise between size and efficiency. What’s interesting about the quantum-inspired approach is that it uses very abstract mathematics to reduce redundancy much more than usual.”