Quantum-Inspired Hack Removes Censorship From DeepSeek R1 and Cuts Model Size by 55%

Summary: Researchers at Spanish quantum-computing firm Multiverse report they have removed built-in censorship from the compact DeepSeek R1 model while compressing it by roughly 55% using a method called CompatifAI. The team says the compressed model shows only minimal accuracy loss and answered previously blocked prompts about events like Tiananmen Square and comparisons involving President Xi Jinping.

Background

Earlier this year, the open-source model DeepSeek R1 drew attention for delivering high-quality responses while requiring substantially less compute than many competitors. Its efficiency even contributed to market turbulence as investors reassessed AI spending. However, R1 strictly followed Chinese censorship rules and declined to respond to a number of politically sensitive prompts.

What Multiverse Says It Did

Researchers at Multiverse say they developed a proprietary compression method called CompatifAI that both prunes the model and removes learned censorship behaviors. According to the team, the technique reduced the model’s parameter count by about 55% while producing only minimal accuracy degradation on standard tests.

How CompatifAI Works

CompatifAI reportedly identifies and removes low-impact parameters—those that contribute little to overall performance—while preserving the core capabilities of the model. The approach uses quantum-inspired mathematical tools known as tensor networks to compress and reorganize large arrays of learned values. In effect, the method selectively prunes learned behaviors, including those associated with censorship, rather than applying a blunt filter or rule set.

Examples and Results

After compression and pruning, the modified model responded to queries that R1 previously refused to answer. For instance, it provided a general analysis of the implications of removing presidential term limits and described historical events previously blocked by the original system. The team reports these changes came with only minimal drops in benchmark accuracy.

Note: Multiverse’s claims are reported by the researchers. Independent verification and wider community testing will be important to confirm effectiveness, safety, and any unintended side effects.

Implications and Concerns

The work highlights how model behaviors—even those that appear normative—can be encoded in parameters and thus altered by targeted pruning. That has broad implications: it may be possible to remove unwanted biases or restrictions, but doing so also raises safety, ethical, and legal questions about who controls model behavior and how.

Experts and analysts caution that removing censorship from a deployed model does not erase the influence of censored or curated training data. If the dataset used to train a model already reflects systematic omissions or distortions, those patterns can persist even after pruning. Additionally, unregulated removal of safeguards could lead to the spread of harmful or manipulated information.

What’s Next

Multiverse’s results, if replicated, demonstrate a promising direction for model efficiency and fine-grained behavior editing. Still, broader evaluation by independent researchers, transparency about methods, and careful consideration of ethical and legal implications will be essential before such techniques are adopted at scale.